“Quality of code (and everything that goes along with it) isn’t talked about enough in AI conversations! There are some obvious facets to this – does the code do what you intended? Is it fast? Does it crash in the common cases?”
Q1. Can AI write better code than humans?
Mark Williamson: I don’t think so, at least not today. For one thing, LLM-based AIs are trained on pre-existing code, which was written by fallible humans. So they at least have the potential to make all the mistakes we do.
Despite that, any coding AI you pick will write better frontend Javascript than me – that’s not my area of expertise. But I would back an experienced human (with or without AI assistance) to beat an unsupervised AI coder.
Can they beat humans some day? I assume so – but they’re not doing it today. And when you factor in other aspects of the Software Engineer’s job (such as building the right thing) it’s even more challenging.
Q2. How do you define what is a “better” code?
Mark Williamson: Quality of code (and everything that goes along with it) isn’t talked about enough in AI conversations! There are some obvious facets to this – does the code do what you intended? Is it fast? Does it crash in the common cases?
A lot of the work a human developer does to achieve this is actually achieved after the initial code is typed in. There’s an iterative process of learning about and refining the solution – understanding what you’ve made and improving on it. A lot of this is really debugging, in the broadest sense of the term: the code doesn’t do what you expected and you need to understand and fix it.
There’s another step beyond that, though – whether the code fits its intended purpose. Getting that fit requires understanding the end user, thinking through the implementation tradeoffs and anticipating future developments. For now, I see AI as freeing up some time so we can create space for those human insights.
Just focusing on how many lines of code we create is a pattern in the industry – we overvalue simply generating code versus all the other things that software engineers actually do.
Q3. Can AI write some types of code faster and with fewer simple errors?
Mark Williamson: Yes!
In my experience, I’ve found AI to be extremely useful in three scenarios:
- Writing code that is almost boilerplate – where it’s not a copy-paste problem but requires quite routine changes.
- Writing code that would be boilerplate for a different engineer – e.g. if I want to write JSON serialisation / deserialisation code in Python it’s easier for me to get an AI assistant to show me the shape of a good solution.
- Doing refactors that involve restructuring or applying a small fix in a lot of places – a coding agent can handle the detail while I concentrate on the overall shape.
In all these cases, the benefit is in reducing the amount of thinking required to figure out my design approach. In Daniel Kahneman’s book Thinking Fast and Slow, he describes two modes of thought: System 1 and System 2. System 1 is the stuff you can just answer automatically, whereas System 2 thought requires effort.
System 2 is tiring – you probably can’t manage more than a couple of hours of really hard thinking about code in a day. So it’s precious. An agent lets me offload some work so I can focus that effort on exploring solutions to the real problem I’m trying to solve.
Q4. Large Language Model (LLM)-based AI code assistants are powerful tools, but they have significant limitations that developers must understand. What are such limitations?
Mark Williamson: The most obvious limitation is that they don’t know everything. They often act as though they do, which is a trap. “Hallucinations” are the most well-known consequence of this – in which the LLM gives an answer that is confident but ultimately not based in fact.
I like to say that modern AI’s training teaches it what a good answer looks like – they’ve seen lots of examples of them, after all. So, from an AI’s point of view, a good answer includes attributes like:
- Projecting confidence.
- Using the right terminology.
- Relating suggestions specifically to your question and context.
- Being right!
If it can satisfy most of those, then it’ll think it’s done a good job. So when they’re asked a question and they lack facts, an AI will figure “3 out of 4 isn’t bad” and give a dangerously convincing answer that’s not based in reality.
There are two important things we can do to reduce this risk:
- Supply high-quality context to the underlying model – the more relevant information available the better. Supplying insufficient information invites the model to guess and supplying irrelevant information encourages it to head off on the wrong track.
- Verify the model’s answers against a ground truth – run your tests, have experts review your code, verify the dynamic behaviour of the application matches what you expected.
You want to focus the model’s intelligence on solving the real problem (not on guessing), then know when it has actually solved it.
Q5. While LLM-based code assistants are incredibly powerful, there is critical information they lack that limits their effectiveness and makes human oversight essential. Why this?
What does it mean in practice?
Mark Williamson: As a CTO, I’ll divide my answer into two parts:
- As an engineer, LLMs don’t know enough about your code to solve all the problems you wish they could solve. They typically don’t have good knowledge of the runtime behaviour of the system, which makes incorrect answers more likely. And they’re not good at inferring design intent, making it harder to fix subtle bugs correctly.
- As a product manager, LLMs lack the insight into the true purpose of the software to be built. You cannot rely on them to design the code to the needs of the end users, long term evolution / maintenance and business tradeoffs required.
Q6. LLMs are brilliant at static analysis—interpreting the text of a codebase, logs, and other documents. But they are blind to dynamic behavior. This is the critical information they lack and cannot get. Why? Do you have a solution for this problem?
Mark Williamson: Coding agents have a similar weakness to humans: they can’t see what the program really did at runtime and it’s hard to reason about why things happened. They can get some of this from logs (and LLMs are really good at reading logs!) but logging can only capture so much.
There’s a catch 22 here for the developer: if you’d been able to predict precisely what logging you’d need to fix the bug you’re investigating, then you’d have known enough to avoid the bug in the first place. There’s no reason to think that’s different for LLMs.
Coding agents can follow the same tedious loop that humans do: adding more logging to a codebase and running stuff again (or perhaps asking a human to obtain more logs some other way).
They can even do this toil more enthusiastically than any human! But the speed you gained from the agent may just disappear into a swamp of rebuilding, attempting to reproduce, finding what logging statements are still missing and then repeating the process. This kind of inefficiency will be bad news for any Engineering department hoping to improve productivity in return for their AI spend.
Q7. It seems that time travel debugging (TTD) directly addresses this limitation. Please tell us more.
Mark Williamson: Time travel debugging captures a trace of everything a program does during execution. The resulting recordings effectively represent the whole state of memory at every machine instruction the program executed.
Anything you want to know about the program’s runtime behaviour can then be queried from the recording, without needing to re-run or change the code. Rare bugs become fully reproducible and any state can be explored in detail. Moreover, the ability to rewind time makes it easy to explore why a bad state arose, not just what the state was.
Of course, storing all of memory at every point in execution time would be extremely inefficient! A modern, scalable time travel debugger stores only information that flows into the program (initial memory state, IO from disk and network, system calls results, non-deterministic CPU instructions, etc). This makes it possible to efficiently recompute all other state on demand. Watch the talk “How do Time Travel Debuggers Work?” for the full details on how a modern time travel debugger is built.
For an AI, this capability is ideal. Remember that we need high-quality context to feed the model and a ground truth to make sure its answers are based in reality. With time travel debugging, a coding agent has access to a recording of the program’s dynamic state and can drill down in detail on any suspicious behaviours – that gives us high-quality context. The ground truth comes from the deterministic nature of the recording and also makes it possible to verify the AI’s findings.
These properties mean that AI coding agents get smarter when given access to a time travel debugging system.
Q8. You have released an add-on extension called explain, which integrates with your UDB debugger (part of the Undo Suite). What is it and what is it useful for?
Mark Williamson: Good question. Let me explain first what Undo is to set the context. It’s our time travel debugging technology (which runs on Linux x86 and ARM64) and is mostly used to debug complex enterprise software that makes use of advanced multithreading techniques, shared memory, direct device accesses, etc.
The Undo Suite captures precise recordings of unmodified programs using just-in-time binary instrumentation.The two main components of the Undo Suite are:
- LiveRecorder – which captures program executions into portable recording files.
- UDB – which provides a GDB-compatible interface to debug both live processes and recordings (but also integrates into IDEs such as VS Code).
The explain extension is our first step in integrating AI with a time travel debugging system. It provides two pieces of functionality:
- An MCP (Model Context Protocol) server – this exports the functionality of our UDB debugger for use by an AI agent, allowing it to integrate into existing AI workflows including agentic IDEs (such as VS Code with Copilot, Cursor or Windsurf).
- The explain command itself, which provides additional tight integration with terminal-based coding agents (such as Claude Code, Amp and Codex CLI) where available.
In either case, we’re providing the power of time travel debugging to an AI, so that it can reason about the dynamic behaviour of a program. As the name suggests, this extension has a particular focus on explaining program behaviour – how a given state arose, why the program crashed, etc.
We provide a carefully-designed set of tools to the agent so that it can answer these questions effectively. It’s important that the design of the MCP tools guides the actions to be taken by the LLM, otherwise it can easily get overwhelmed by the complexity.
In an agentic IDE you can connect to the MCP server in a running UDB session – then ask the agent questions (use the /explain prompt exported by the server for best results). In UDB itself, you can just type the explain command and we’ll automatically invoke your preferred terminal coding agent and put it to work on your problem.
Q9. Can you show us an example of how time traveling with an AI code assistant works in practice?
Mark Williamson: Sure! I’d recommend watching these two demo videos:
- The cache_calculate demo video on the Undo website which showcases how to use explain to get AI to tell you what has gone wrong in the program.
- This YouTube video where I use AI + time travel debugging to explore the codebase of the legendary Doom game and understand exactly what the program did when I played it.
We have additional demos, showcasing more advanced functionality, which aren’t yet public – you can book a personalised demo from https://undo.io/products/undo-ai/ to see the more advanced AI debugging functionality we’re currently building.
Qx. Anything else you wish to add?
Mark Williamson: The core message here is that AI-Augmented Software Engineers still need the right tools to do their jobs well. Our goal is to make AI coding agents more effective at understanding and fixing complex code, improving the return on investment Engineering departments get on their AI stack.
The next big step for us will be designing a UX to be used by AIs instead of by humans. Providing time travel debugging to a coding agent is already useful, but to get the best performance we need to work with what LLMs are good at. In other words:
- A query-like interface: rather than the statefulness of a debugger, LLMs are happiest when they can ask Big Questions and get a report in answer. Our engine lets us extract detailed information very quickly from a recording so that an AI can start with an overview, then drill down.
- Specialised, composable tools: a debugger provides quite general tools (stepping, breakpoints, etc) for a human developer to apply to any problem. Coding agents can use these but we believe LLM intelligence is best spent on solving the core problem well, rather than diluting it on planning complex tool use. A specialised set of analyses will allow the LLM to focus on what it’s good at – finding patterns and proposing fixes.
On top of these tools and the data contained within our recordings, we are building Undo AI – a product to enable agentic debugging at enterprise scale. We’re currently taking applications for our pilot program, please get in touch to find out more at undo.io .
……………………………………………
Mark Williamson, Chief Technical Officer, Undo
After a few years as our Chief Software Architect, Mark is now acting as Undo’s CTO. Mark loves developing new technology and getting it to people who can benefit. He is a specialist in kernel-level, low-level Linux, embedded development with a wide experience in cross-disciplinary engineering.
In his previous role, his remit was to align the product’s architecture with the company’s needs, provide technical and design leadership, and lead internal quality work. One of his proudest achievements is his quest towards an all-green test suite!
As Undo’s CTO, Mark’s primary responsibility is to scale product-market fit and ensure we take our products in the right direction to meet the needs of a broader spectrum of customers.
Mark is also author on Medium, a conference speaker, and a new home owner enjoying the delights of emergency home repairs!
………………………..
“I wish more organizations realized how fundamental it is to lay a good foundation for any enterprise AI initiative. That foundation includes a robust data strategy and a unified architecture.”
Q1. What are the responsibilities of a Field CTO?
Stephen Kallianos: As a Field CTO, my core responsibility is to serve as a trusted advisor for enterprise customers, especially at the Senior Architect and C-level. It’s a consulting role, focused on establishing credibility and helping customers and prospects connect their strategic priorities to SingleStore’s unique value proposition.
The work involves identifying and clearly communicating the best-fit enterprise architectures, leveraging deep expertise in data and AI infrastructure. My role requires me to thoroughly understand customer challenges, align our technical solutions to those needs, and recommend the most effective solutions.
In addition, I lead the presales function here at SingleStore: running technical discovery, developing tailored demonstrations and proofs of value, qualifying opportunities, and shaping value-based engagements that bridge the gap between technology and business results. Ultimately, my goal is to ensure that our solutions deliver both technical and business impact – setting organizations up for long-term success in their modernization efforts.
Q2. What’s something you often hear from customers and prospects?
Stephen Kallianos: Organizations are hungry for applications that can leverage the most recent data for AI-driven insights, but often get bogged down managing separate systems for transactional and analytics workloads — leading to increased complexity and database sprawl. I regularly hear concerns about inconsistent query performance, missed SLAs for real-time or batch data, and the growing need for flexible deployment options — be it cloud, on-prem, or hybrid. Most notably, there’s a surge in organizations looking to modernize: they want to drive real business outcomes by reducing operational overhead, simplifying their technology stacks, and future-proofing their data infrastructure to keep pace with rapidly evolving AI requirements and new digital experiences.
Q3. What are the most commonly shared pain points among customers seeking to implement enterprise AI?
Stephen Kallianos: Customers implementing enterprise AI encounter a few pervasive pain points. Common issues include: navigating data silos and complex integrations, struggling to perform large-scale aggregations efficiently, and dealing with the high costs and poor performance that come with scaling legacy data infrastructure. Meeting real-time data requirements for AI workloads is a particular challenge, especially when data resides in multiple, disparate databases. Legacy architectures often fail to deliver the query performance and SLAs necessary for AI use cases, leading to a pressing need to modernize and consolidate systems.
Q4. Do you see GenAI being used in the enterprise? How?
Stephen Kallianos: Absolutely. Enterprises are rapidly adopting generative AI (GenAI). They’re integrating large language models (LLMs) into their AI architectures for a range of scenarios from analytics and customer support to productivity tools and operations. We’re seeing production deployments in areas like enterprise search (retrieving contextually relevant records and documents), AI-powered personal assistants and co-pilots, workflow automation, developer productivity tools (text-to-SQL, code recommendations), and even advanced analytics for fraud detection and data enrichment
Q5. What do you wish more organizations knew when it comes to adopting enterprise AI?
Stephen Kallianos: I wish more organizations realized how fundamental it is to lay a good foundation for any enterprise AI initiative. That foundation includes a robust data strategy and a unified architecture.
Relying on siloed or hastily patched-together systems makes it almost impossible to achieve the simplicity, security, or scale needed for AI to succeed in production. The best results come from adopting a single platform that handles analytics, machine learning, and operational workloads — streamlining architecture and lowering risk. Ultimately, AI projects succeed when technical outcomes are tightly aligned to clear business value — not when technology is adopted for its own sake.
Q6. How can organizations better align their technical solutions with the organizational goals?
Stephen Kallianos: In my experience, the best way organizations can align technical solutions with their business goals is for the organization and the database vendor to hold a workshop to clarify and align on both the desired business outcomes and technical requirements. Mutually qualifying opportunities up front — ensuring there’s clarity and genuine need — helps avoid wasted effort later. Together, you can frame what success looks like and define concrete criteria, creating a North Star for architecture, implementation, and measuring results. Having hands-on proof-of-value phases (using real data and involving cross-functional teams) is key to validating that proposed solutions actually deliver the anticipated outcomes, and is an approach I use extensively when leading presales and customer workshops.
Q7. SingleStore recently unveiled a new version of the database, and it contains a lot of upgrades. In your opinion, which 2-3 things are most valuable to customers? Why?
Stephen Kallianos: For me, several features from the latest SingleStore release stand out as particularly remarkable. These include the major upgrades we made to Flow, Iceberg, Aura, and our developer experience.
The first is about ingest and data integration. With SingleStore Flow (our no-code solution for data migration and continuous change data capture) now natively embedded in our Helios managed service, customers can orchestrate data movement into SingleStore directly within the cloud platform, making the process far more streamlined. Data ingestion is now much simpler and more flexible, and moving data from heterogeneous sources like Snowflake, Postgres, SQL Server, Oracle, and MySQL is easier than ever.
This is part of our overall “SingleConnect” experience that allows customers to incorporate more and richer data sources into SingleStore. Adding Flow into SingleStore Helios® further strengthens our ability to integrate from diverse environments, reducing integration friction and enabling real-time analytics and AI use cases without the pain of traditional ETL complexities.
We’ve also done a lot to enhance our Apache Iceberg ecosystem. For customers using data lakehouses with Apache Iceberg, there’s now a speed layer that offers high-performance, low-latency data interaction on top of Iceberg-managed storage. Improved bi-directional integration allows for easier, faster data exchange with external Iceberg tables, so real-time applications can finally tap into lakehouse architectures with the latency and interactivity they require.
In the area of AI and serverless compute, we upgraded our Aura Container Service. Aura brings together vector search, analytics, function-as-a-service, and GPU-accelerated workloads in a single containerized environment. Already optimized for running AI/ML and containerized workloads, Aura now offers support for cloud functions (lambda-style serverless functions). This unlocks the ability to build data APIs, agents, and inference endpoints for embeddings or other ML tasks, all within a managed, scalable environment. Coupled with performance enhancements like multi-value indexes for JSON, automatic query re-optimization, and improved cross-workspace branching and disaster recovery, these upgrades drive higher reliability and enterprise scalability.
We’re always thinking about the people doing the work, so we made substantial improvements to the developer experience, with enhancements to our AI-powered query builder assistant (SQrL), deeper integrations with GitHub, notebook scheduling/versioning, better pipeline and billing visibility, and a more powerful multi-tab SQL editor. All of these improvements make building, monitoring, and scaling AI and data applications faster and more seamless.
Collectively, these advances eliminate bottlenecks, simplify integration, and provide the speed, flexibility, and full-lifecycle support today’s enterprise AI and analytics apps demand.
Q8. Which is more prominent, on prem or cloud? With security and privacy being big concerns these days, are people talking about returning on-prem?
Stephen Kallianos: Our strategic advantage is that we are truly hybrid — with the most versatile offering across SaaS (Helios), Bring Your Own Cloud (BYOC), and self-managed solutions. We provide maximum flexibility for customers to deploy anywhere, on any cloud, allowing them to meet their specific business, technical, and regulatory needs. We’re seeing continued momentum around our managed cloud service, Helios—driven by a desire for operational simplicity, scalability, and innovation. But our BYOC and self-managed (private cloud) solutions are also going strong. This flexibility means customers can mix and match approaches: leveraging Helios for fully managed simplicity, BYOC for deployment control, or self-managed options for maximum security and privacy. Ultimately, we empower customers to modernize on their terms, run workloads wherever they need, and never have to compromise on control, compliance, or agility.
Q9. What’s next for the industry, and what is SingleStore doing to help meet those needs?
Stephen Kallianos: The next evolution in our industry is all about convergence — bringing together analytical, transactional, and AI workloads on unified data platforms. Today, less than 1% of enterprise data is being used for enterprise AI, so the opportunity is immense. There’s a heightened focus on delivering real-time intelligence, integrating AI natively, and eliminating data silos, alongside surging demand for seamless integration with data lakes, warehouses, and GenAI/LLM platforms. SingleStore is innovating aggressively by expanding serverless compute, adding integrated AI and ML functions, launching AI co-pilots, enabling direct LLM integration, and introducing the Aura platform that I mentioned earlier. These advances are designed to enable customers to build the next generation of data-driven and AI-powered applications — unlocking more value from their data and making enterprise AI real for the business.
………………………………………
Stephen Kallianos, Americas Field CTO, SingleStore.
With deep expertise in data-driven strategies and cloud-based innovation, Stephen Kallianos is the Americas Field CTO at SingleStore. In this role, he combines his SingleStore expertise and industry knowledge to drive a collaborative approach towards helping customers align solutions with business goals.
………………………..
“ Trades are driven by real-time market conditions where billions of dollars move every second, generating enormous amounts of data. The biggest challenge is minimizing the latency associated with analyzing these chaotic data streams and turning it into something that’s actionable for traders.”
Q1. What is your role at ExeQution Analytics?
Cat Turley: I’m the CEO and founder of ExeQution Analytics. We’re a boutique consultancy focused on helping financial organizations, particularly trading firms, take greater advantage of their data infrastructure. The story of ExeQution Analytics began 20 years ago, when I was working at an international broker. I challenged the head of trading to “do more” as I believed that we could write more interesting analytics and achieve better understanding of the markets and our trading patterns. I truly believed we were only skimming the surface of what kdb+ could achieve. He returned the challenge and invited me to build a green-field analytics platform capable of understanding market microstructure and providing real-time and historical signals to electronic trading strategies.
Over the past two decades, I’ve continued to refine this approach to analytics. Four years ago, we officially launched ExeQution Analytics as demand had grown, and we identified a gap in the market. There were plenty of resources focusing on the acquisition and storage of data, but less focus on what the data was used for. We developed a structured and flexible analytics framework to solve the problem that everyone was seeking to solve: how to make analytics more efficient and accessible across all aspects of the trading lifecycle. Now my role requires that I work closely with those we have partnered with, from financial organisations on both sides of the street, to technology leaders such as KX.
Q2. How do you help organizations maximize the value of their technology investments and improve data-driven innovation?
Cat Turley: What makes ExeQution Analytics unique is that we’re positioned right at the intersection of the three pillars of trading: the traders themselves, quants and technology leaders. We speak all three languages and provide a framework that helps everyone achieve their common goal of delivering better trading outcomes. Our standardized framework efficiently analyzes large volumes of market data at speed and scale. From there we create customized analytic platforms that enable firms to gain enhanced and actionable insights tailored to their unique trading workflows.
Trading teams cannot accelerate innovation if they’re stuck spending all of their time preparing data. We’re giving them the tools necessary to remove the onus of data preparation and instead focus on extracting signals, identifying patterns, and understanding market activity. When armed with these insights, firms can test more ideas, move faster, ask better questions about their data, and ultimately generate strategies that improve trading outcomes.
Q3. Let’s talk about Quant. What do they do and how have they evolved?
Cat Turley: Quant teams build and refine models that power trading strategies, everything from price prediction to portfolio optimization. This has always been a data-driven process, but over the years, thanks to the advancement and accessibility of computational tools, it has increased in complexity and sophistication. Now, quants can efficiently and quickly analyze years of historical market data to glean unique insights that optimize market prediction and trade execution.
Most financial organisations have been using advanced machine learning capabilities for the last decade or so, enabling more sophisticated predictions. There is potential for even further evolution as advances in AI become more integrated into the quant trading process through the use of large language models, vector databases and techniques such as time series similarity search.
The second significant avenue of evolution is the integration of real time market data into the quant lifecycle, enabling better understanding of how models react to the ever-evolving market conditions. As data volumes grow, it has never been so important to remain agile in volatile market conditions.
Q4. If we consider Intra Trade Monitoring: What are the challenges?
Cat Turley: Trades are driven by real-time market conditions where billions of dollars move every second, generating enormous amounts of data. The biggest challenge is minimizing the latency associated with analyzing these chaotic data streams and turning it into something that’s actionable for traders. Trading once relied heavily on human intuition and experience with many decisions based on “gut feeling”, but with the advances in markets and technology, this instinct can now be augmented with data-driven understanding. The challenge is getting the right analytics in front of the right person at the right time, so they can make the best decision to improve trading outcomes. These days, traders are typically monitoring thousands of individual orders at any one time, as algorithms control execution. They need access to tools that can distill all this noise into actionable insight.
Q5. How is Trading Analytics related to Quant and Intra Trade Monitoring? What do you see as major challenges here?
Cat Turley: Intra trade monitoring supports real-time observation and analysis of trade execution, feeding live data into analytics systems. Both traders and quant analysts depend on these analytics to measure performance feedback for refining computational models that drive pricing, forecasting, and trade decision-making. Essentially, these three components support pre- and post-trade analysis.
The challenge many firms are facing is how to use trading analytics to transform TCA from a tick-the-box exercise to a more comprehensive framework to properly understand the nuances and intricacies of trading execution and opportunities for optimisation. Historically, pre-trade and post-trade were often considered separate processes. To truly optimise execution, they should consider two aspects of the same process, where one informs the other.
This is where KX can offer an advantage: one of its unique attributes is that it’s a high-performance analytics database optimized for both real-time and historical data. Building a custom trading analytics platform using KX technology allows organisations to evolve towards more proactive analytics, enabling the identification of both optimisation opportunities as well as execution risks and alpha generation opportunities. Integrating real time data into the TCA/execution analysis or trading research process enables a better understanding of where a different trading decision would have resulted in an improved outcome, and back-testing with historical data can inform where this intuition offers statistically significant performance improvement.
Q6. You have been using KX for over 20 years now. How did the KX ecosystem evolve over time?
Cat Turley: It’s changed dramatically. When I started over 20 years ago, obtaining a kdb+ license required a much more substantial investment in development resources, often turning into a one-to-two-year project before it was put into production. Now, thanks to KX-driven platform releases and updates as well as tools like Data Intellect’s TorQ, the development of kdb+ infrastructure has been streamlined. Today, teams can take advantage of previous iterations and go to market much faster. It’s gone from a highly technical, custom-built process to something far more streamlined, accessible, and scalable. That means firms can focus on the nuance of their individual trading requirements, how to turn data into value, rather than spending as long on the building blocks of data ingest, storage and availability.
Q7. You use kdb+ for trading. What are the main benefits you see in using such a database?
Cat Turley: ExeQution Analytics framework is developed in q and designed for integration with kdb+ platforms. So, you could say that I’m a big advocate for q and kdb+, and that’s not just because of its unprecedented speed—but also the incredible flexibility that the q query language offers – it truly is the standout benefit. kdb+ is the fastest time-series database available, and q enables us to move beyond a data-warehouse to deliver genuine analytic platforms with a reduced time to market. That speed allows quants to pursue excellence – they can fail fast, learn fast, and keep improving their models.
As I briefly mentioned earlier, kdb+ is unique because it can handle both real-time and historical data without compromising speed and performance. In the trading world, this combined benefit is what allows for better trading outcomes, rather than a series of missed opportunities.
Q8. Specifically, how do you handle large volumes of real-time and historical data with low latency? Why does kdb+ being columnar matter?
Cat Turley: Our approach combines fast in-memory processing for live data with efficient on-disk, columnar storage for historical data, enabling seamless and high-speed time-series analytics across both. And kdb+’s columnar absolutely makes a difference in optimizing low-latency performance because it only pulls the fields needed.
Storing data in columns allows for faster reads, better compression, efficient CPU caching and parallel processing, all of which are ideal for fast-moving, analytical trading workloads.
Q9. kdb+ offers Q, a SQL-like language. How easy is it to use, and how do you encourage adoption over SQL?
Cat Turley: I am a huge advocate for the q language – it is elegant, expressive and enables incredibly fast time-to-market from a development perspective. An analytic written in q versus SQL or Python is typically 10 times more concise. This means it takes 10% of the time to write, your development team can be 10% of the size, and you incur 10% of the errors. It may have a reputation for being harder to learn but it is well worth getting over the initial learning curve, as it offers huge benefits once mastered, especially when dealing with time-series operations, high-frequency data and real-time decision making.
The other benefit is that the q community is unlike any other; you can really lean on these folks for support and learning materials. SQL gives you access to data and is well-suited for general purpose tasks and projects, but q is purpose-built for speed and analytics at scale which are critical benefits for those working with high-speed or high-volume data.
Q10. kdb+ is used across hedge funds, investment banks, and trading firms. What are the similarities and differences among them when dealing with quantitative trading operations?
Cat Turley: Their trading operations are similar in the sense that they are all working with high data volumes and require low latency. Of course, they have varying levels of latency tolerance and data flow, but overall, I would say the premise is the same: operations need to optimize data-intensive workloads and minimize time-to-decision.
What I think differs is their objectives. Hedge funds prioritize strategy simulation and alpha generation, banks emphasize client service and pricing, and trading firms are laser focused on speed and execution edge. Regardless of the objective, all rely on the ability to process massive volumes of real-time and historical data with precision and speed.
………………………………………………………………
Cat Turley, CEO/Founder, ExeQution Analytics
With 20 years’ experience working with leading global investment banks and some of the world’s largest asset managers, Cat has an extensive understanding of market microstructures, execution analysis and how the right choice of technology can empower organisations to achieve more with less. Cat is passionate about improving efficiency and understanding across all areas of trading and founded ExeQution Analytics to contribute towards this goal.
Related Posts
On Trading Tech and Quant Development. Interview with Jad Sarmo, ODBMS Industry Watch, May 29, 2025
………………………..
Forecasting financial time series is one of the most complex tasks in data science.
Q1. You’ve been working in the Trading Tech and Quant Development space for the last 20+ years. What are the main lessons you’ve learned through this experience?
Jad Sarmo: Back in 2004, I deployed the first automated trading system (ATS) for foreign exchange at a top-tier bank. We had to build software directly on traders’ workstations to send algorithmic orders—latency was measured in hundreds of milliseconds.
Since then, the landscape has evolved dramatically: the proliferation of low-latency submarine fiber-optic cables, high-frequency signals bouncing off the ionosphere, the emergence of cloud computing, AI-assisted development, the rise of blockchain, and nanosecond-level FPGAs.
Despite this, the core principles remain unchanged: a solid grasp of systems and markets, clear business objectives, and the ability to assemble the right experts to solve the right problems. Equally important—especially as firms face increasing external scrutiny and apply for new licences—is a commitment to compliance with applicable laws and regulations from the outset.
A personal lesson I’ve come to value is this: if you’re comfortable, it’s time to take a risk, learn, and repeat. That cycle is essential in such a fast-evolving landscape.
Q2. What is your role at B2C2?
Jad Sarmo: B2C2 are a global leader in the institutional trading of digital assets, serving institutions such as retail brokers, exchanges, banks, and fund managers. We provide clients and the market with deep, reliable pricing across all market conditions.
I joined B2C2 in 2021—during a pivotal year for digital assets—to build our global Quantitative Development desk. My team works closely with traders, researchers, and engineers to improve client pricing, trading strategies, and automated risk systems.
I also lead our Exchange Squad, which manages trading from market data ingestion to algorithm optimization across more than 30 AWS regions globally.
Q3. What are the main challenges in this industry when it comes to data management? Specifically, since you are handling liquid assets, what is the main challenge you’ve seen when an asset can be “easily” converted into cash in a short amount of time?
Jad Sarmo: Like in any asset class—FX, equities, rates—crypto trading involves massive volumes of market and trading data. But crypto adds a unique layer of complexity.
It’s a 24/7 market, with both on-chain (blockchain-logged) and off-chain (centralized exchanges) activity. A significant share of volume also flows through DeFi protocols using smart contracts.
We face challenges like inconsistent exchange APIs (REST, WebSocket, etc.), cloud-native environments, and the need for extremely low-latency systems that handle massive data bursts. Meanwhile, newer or illiquid tokens present formatting hurdles, with decimals occasionally extending to 10+ digits — far beyond what many traditional systems were designed to handle.
Real-time hydration and normalization of incoming data streams are therefore critical to support both research and trading effectively.
Q4. You mentioned in a previous presentation that managing a “Crypto ecosystem” is not an easy task. What is a Crypto ecosystem, and what is it useful for? What are the specific challenges you face, and how do you solve them?
Jad Sarmo: By “crypto ecosystem,” I mean the global, interconnected infrastructure where digital assets are traded: exchanges, OTC counterparties, and all supporting systems.
Each participant may be located in a different —Virginia, Tokyo, London, and beyond. Our system ingests high-frequency data from across the world, unifies it, and processes it with both low latency and high throughput.
The hardest part is normalizing inconsistent feeds so they’re useful across trading and research. Historically, AWS prioritized reliability over low latency, but in recent years, the biggest players—including B2C2—have worked closely with AWS to re-architect the cloud to meet the latency needs of crypto trading.
Q5. Let’s talk about the use of AI and Machine Learning in the financial services industry. You cannot predict the market by training an AI model on historical data, because things change rapidly in the financial markets. How do you handle this issue? Does it make sense to use AI?
Jad Sarmo: Forecasting financial time series is one of the most complex tasks in data science.
A picture of a dog from 10 years ago is still useful to train an image classifier—but financial data ages fast. Market structure, participants, and behaviour shift constantly, so models need regular recalibration.
Ensemble learning is particularly powerful in finance; rather than relying on a single predictive model, we combine many models that each perform slightly better than average. AI is not a crystal ball, but it provides meaningful signals that enhance traditional pricing and risk systems.
Q6. You have been leveraging a vector-native data platform at B2C2. Could you please explain what you do with such a data platform?
Jad Sarmo: We use KX’s kdb+ platform to support our real-time and historical time-series data needs. It enables global ingestion across AWS regions, persistent storage, replay of massive tick datasets, and complex event processing.
The consistency of this platform means researchers can focus on analysis without worrying about where the data lives. PyKX, a Python–Q hybrid notebook interface, allows heavy computations to run in Q, while using Python for exploratory analysis and ML.
KX also provides high-performance dashboards for quick data visualization—even by non-technical users.
Q7. Why not use a classical relational database or a key-value data store instead?
Jad Sarmo: Traditional relational databases are too rigid and slow for high-frequency time-series analytics. Key-value stores are great for quick lookups but lack native analytics support.
Vector-native platforms like kdb+ are designed for exactly this use case. They let us run complex queries over billions of rows in milliseconds—without reshaping the data or creating indexes.
As data volume grows to terabytes per day, traditional databases become engineering bottlenecks. In contrast, vector platforms scale naturally, with each column and date efficiently mapped to files.
Q8. Let’s go a bit deeper. If you start with “FeedHandlers,” how do you end up processing this complex data at scale, in real time and without losing some data?
Jad Sarmo: Our architecture begins with Java or Rust feed handlers that convert raw exchange data into kdb+ format.
A ticker plant then routes data to three layers:
1 A real-time in-memory database
2 A persistent on-disk database
3 A complex event processor
This setup ensures we can act on data instantly, store it reliably, and support deep analytics—all with complete transparency for end users, whether they’re consuming live or historical data.
Q9. What about data quality? How do you ensure data quality in the various phases of data processing?
Jad Sarmo: Data quality starts with ingestion. Exchange feeds vary in reliability and format, so we normalize and hydrate the data immediately to remove inconsistencies.
We maintain constant feedback loops between research and production teams to monitor and improve quality. Clean, consistent data is the backbone of everything—without it, even the most sophisticated models won’t perform.
Q10. You decided to integrate AWS FSx for Lustre with kdb+. What are the main benefits of this design choice?
Jad Sarmo: AWS FSx for Lustre has been a major improvement. It offers virtually unlimited horizontal scaling and high-speed access. We can connect dozens or hundreds of nodes, each with fast local disk and compute, to form a massive high-performance network file system.
It compresses files efficiently, offloading that work from kdb+. We can spin up isolated research environments on demand without affecting production, and there’s no downtime. Auto-scaling lets us right-size our infrastructure at any time.
Compare that to traditional datacentres—provisioning takes weeks and usually leads to overbuying hardware. In the cloud, it’s a five-minute job.
Q11. How is industry regulation affecting this complex data management?
Jad Sarmo: Regulation is advancing quickly. This means we must store data in auditable, retrievable formats. End-to-end traceability—from ingestion to storage to downstream consumption—is non-negotiable.
This adds operational overhead, but it also emphasizes the need for trustworthy systems that meet both performance and compliance standards. We see this reflected in regulatory initiatives like the EU’s MiCA regulation, the approval of Bitcoin ETFs in the U.S., and the UK’s FCA Discussion Paper DP25/1, which explores regulating crypto asset activities.
……………………………………………………………
Jad Sarmo, Head of Quantitative Development | Expert in High-Performance Trading Systems, B2C2.
Jad Sarmo is a technology and trading infrastructure leader with over 20 years of experience building high-performance trading systems for FX and digital asset markets. He is currently Head of Quantitative Development at B2C2, a global leader in institutional liquidity for digital assets, where heoversees a global team delivering real-time pricing, exchange trading, and analytics infrastructure across 24/7 markets.
Prior to B2C2, Jad ran Technology at Dsquare Trading, a high-frequency proprietary FX trading firm that rose to prominence through cutting edge algorithms, low-latency engineering, and a world-class team. There, he designed ultra-fast trading systems and led cross-functional teams through years of continuous innovation in a high-stakes environment.
Jad specialises in bridging the gap between trading, quant research, and engineering — turning complex ideas into reliable, automated, and profitable systems. His expertise spans real-time architecture, algorithmic trading, market data, and risk management, with deep technical fluency in Java, Python, KDB+/q, and AWS.
Based in London, Jad is dedicated to designing robust systems under real-world constraints and mentoring the next generation of technologists.
………………………..
Synonymous with innovation, R&D is all about competitive edge. Generative AI has emerged as an R&D catalyst, promising to unlock new use cases and unexplored opportunities to increase margins without extra spending.
We recently assembled an expert panel of seasoned scientists and information research leaders from the Dow Chemical Company to discuss how they built a semantic data hub with the Progress Data Platform to capture years of R&D knowledge, facilitate its discovery across the organization and make it future-ready for years to come.
This post is a summary of the information presented by the panel, which did a much deeper dive into the topics covered. If you would like to watch the interview in its entirety, you can view ithere.
The Panel
From Dow Chemical: Dow is one of the world’s leading materials science companies, with sales of $43 billion (USD) in 2024.
Simon Cook, P.H.D, Senior Solution Manager/Scientist
Alix Schmidt, Senior Data Scientist, R&D Model Deployment Strategy
John Talbert, Fellow and Systems Architect
From Progress:Progress is the trusted provider of AI-powered digital experience and infrastructure software.
Drew Wanczowski, Moderator, Senior Principal Solution Engineer
The material from the Q&A session has been edited for length and clarity.
How does information research support R&D at Dow, and how do you collaborate across business units and R&D groups to navigate complex data and security requirements?
Alix: We act as translators helping scientists capture, store and transform knowledge and data as well as protect it. This is a complex area that requires agility because things in research change all the time. We need to design our data systems so when you find something new or you’re working with something new, we can take that data and be able to incorporate it within our ecosystem very fast.
John: Keep in mind that the scale of Dow and our R&D function includes thousands of people going to the lab, doing experiments, collecting information, documenting information, accessing external information—all kinds of things to innovate and serve our customers and so it’s a pretty large-scale initiative. At the same time, we’re working towards digital transformation to really get to that next level of what does R&D look like in the future and how do we compete with newer digital native companies?
Describe R&D data and how it is different than typical manufacturing data.
Alix: I’ve had about 12 years of experience within the R&D function and five years within manufacturing. When I think of manufacturing data, I think of it as the original Big Data before web analytics. It’s very consistent, the schema stays the same, so you have the same columns, the same tags and other kinds of information within your database.
In contrast, the role of R&D is always to be innovating and doing something new. There are always new pieces of data you’re collecting, new instruments that you might be integrating and variables you might change that you had never changed before. So now you’ve got a different metadata kind of situation. I think the big difference is the agility and being able to just manage those data flows.
What are some of the challenges you see around data silos and data discovery when you have all these varying shapes?
John: Probably one of the big challenges is around consistency across silos as some of the nomenclature may not be consistent from one silo to another. I think that master data is also key if you’re attempting to bring data together from multiple systems. If you are not capturing your identifiers in a consistent way, it makes it very hard to make sense of that data over time. It actually starts in teaching folks that are generating data to understand the importance of capturing metadata as much as possible at the source.
Can you elaborate on your foundational data management approach, particularly around the capture, standardization and discovery?
Simon: If we look at foundational data management, it’s not a static platform. Data systems are constantly evolving so what that really means for any kind of foundational data management is you need to move with those changes and some of the more recent external standards. We have to change the mindset from “I have to pick a standard and stick with it,” to “whatever standard we use, we can change and evolve over time as we need to.” You have to transform data into a form that you can actually use. A lot of our work in that foundational platform involves transforming it into a vendor agnostic format using the best external standards that are out there and then making sure that we can transform them as they change over time.
Can you elaborate a little bit on your approach and the role a semantic data hub plays in tying all this data together?
John: The approach is basically to bring data together and connect it so that people can more easily access it. As we’re aggregating data, we have to understand what’s the context of those data: how should we unify it, should we unify it depending on the use case, etc. If you’re taking data from different systems and just putting it into one place, you haven’t done anything; you haven’t made it any easier to use, you just moved it to a central location. If you want to be able to get the most from those data, master data management is really key. You need to ensure that those pieces are in place as part of your data strategy to ensure that you have the identifiers that are needed when you bring the data together from multiple locations.
How do you make data future-ready rather than future-proof?
Simon: I don’t think there’s such a thing as future proof because we don’t know what the future’s going to bring, apart from our data is going to get more complex. We really need to organize our data the best way we can with our current tools. But whenever we do that, we need to make sure that our data is organized so that we can transform it to anything that’s coming in. In the future, as you enhance existing information with more data, make sure that those structures and concepts you start to build are extensible. Future ready is a better term because it means that you can transform your data rather than going down a blind alley where you’re bought into a standard that’s not translatable to modern data.
How is Dow leveraging GenAI to advance research and take advantage of this new technology?
Alix: We have leveraged both traditional AI machine learning and now generative AI. With GenAI, we’re still trying to give researchers an easy way to access Dow’s extensive internal know-how as well as external information like patents and papers. We think that GenAI along with traditional semantic searching of data can surface some of the context that’s not yet structured. We’re using GenAI to make that information discoverable and usable to support all the other information that our researchers use to make their decisions.
For example, if you wanted to understand an opportunity to make a new type of plastic, you could look through all of the patent literature about plastic and find out what hasn’t been done yet. But that is way too prohibitively time consuming. With GenAI, now there’s ways to process that information to suggest new pathways. The key here is acceleration and the ability to use that massive trove of data within Dow.
Can you take us through some of your requirements for success in setting the foundation up for GenAI?
Alix: At the very beginning, it’s about how do we architect for scaling versus the need to scale immediately. For a GenAI solution to work effectively, we need to focus on quality, interoperable data as the fuel. This entails marrying aggregation with context; standardizing where you can or at least dynamically understand where things are semantically equivalent and where you might have to bring in ontologies and taxonomies that help to describe those relationships.
Can you elaborate on your security requirements and why it is so important to have those in place?
Simon: The number of years of data we have is incredible. If we left that wide open to the external world, we wouldn’t be in business very long. This is why the protection of our data from outside attacks is essential—and that means that we need to put in place many different layers of protection.
Internally, we’re also looking at how do we use the least privilege principle in general—do you need access to all the data, or do you really need access to a subset of the data? Additionally, we’re working with external partners and are very careful to make sure that information between partners is not accidentally leaked across boundaries.
How would you say Progress has supported you and empowered you to build these solutions?
John: We initially brought MarkLogic in to help with our problem with searching literature. Our prior solution was architected for document management, so it did a very good job with the workflow and sign offs.
Now we’re utilizing MarkLogic to help manage literature and enable our researchers to access and find the information they need. We’re expanding that use case to build into new systems such as a data hub where we can aggregate structured and unstructured data from various systems. The data hub is essential for our group’s effort to improve the metadata that’s missing for both making data easy to access and normalizing the security around the data access. As we start to enrich those data, it becomes a really good source for feeding into GenAI models and other types of technologies.
The entire space of how you produce enterprise GenAI solutions is a field that’s only a couple years old, but a few architectures have become very popular, such as RAG, and specifically Vector RAG. The capabilities provided by Progress around knowledge graphs, graph RAG and other types of semantic search are equally as interesting in terms of providing a user with the response they’re truly interested in. We’re not just putting all of our eggs in one basket with Vector RAG—we’re exploring different options and how we can combine the different tools in our toolkit to surface the right information through a hybrid approach.
Watch the Panel Discussion
This post is a summary of the information presented by the panel, which did a much deeper dive into the topics covered. If you would like to watch the interview in its entirety, you can view it here.
………………………..
On Transforming Manufacturing with IoT and Real-Time Data. Interview with Dheeraj Remella.
“We’re also seeing that the latency involved in going to the cloud is becoming a nuisance and there is a lot of movement to bring intelligence to the edge. In this case, edge doesn’t mean “on the device” because that becomes too narrow a context. The edge is a compute location close to all the components and systems in a process.”
Q1. How has the integration of IoT sensors and real-time data processing transformed traditional manufacturing processes since 2020?
Dheeraj Remella: Since 2020, we’ve seen a significant shift. Manufacturers are moving from reactive to proactive operations. Some of the examples are:
- Predictive maintenance is becoming mainstream
- AI-driven decision-making using ML models for real-time inference
- Ubiquitous connectivity through 5G, eSIM, and LoRaWAN has given rise to more data being collected, giving manufacturers immediate visibility into the current state of assets and warehouse parts for faster issue resolution.
- By proactively managing assets and operations, organizations are enjoying energy savings of around 30%
- Digital Twins are being used in the “industrial metaverse” to ensure that a change would have the intended effect before it is implemented in the real world – i.e. the physical twin world.
We’re also seeing that the latency involved in going to the cloud is becoming a nuisance and there is a lot of movement to bring intelligence to the edge. In this case, edge doesn’t mean “on the device” because that becomes too narrow a context. The edge is a compute location close to all the components and systems in a process. So, the organizations need to go beyond the traditional computation of just the equipment efficiency and instead focus on the more meaningful process efficiency.
Q2. What are the key challenges in managing and analyzing the massive scale of IoT data in real-time for smart manufacturing applications?
Dheeraj Remella: Typically, organizations buy software that has its own data storage, which creates silos between processes and departments. These silos create the “Curse of Babel” problem, where every department and software has its own nomenclature and representation. The lack of a commonly accepted and understood ontology creates a significant bump to overcome before true innovation can be achieved. Additionally, the sheer volume, velocity, and variety of IoT data present significant hurdles. Traditional systems struggle to keep up.
Q3. How are edge computing and cloud-based analytics being leveraged to address latency issues in real-time data processing for manufacturing environments?
Dheeraj Remella: Data value is primarily determined by how fresh it is. Younger data is suitable for faster contextual decisions, while older data is better used as a part of a collective for extracting learnings from the system behavior. Naturally, the younger data needs to be acted upon at the edge and the older data should be collected in the cloud for economies of scale for the non-time-sensitive machine learning initiatives.
In addition, edge computing gives the ability to incorporate various secondary concerns into the data processing:
- Security
- Sustainability
- Sovereignty
- Sessionization and aggregation
- Data thinning to send less and just the relevant data to the cloud.
- Digital twins being activated beyond being just data stores allows intelligent participation in the physical processes
- Elimination of redundant or insignificant ephemeral data
Q4. Can you discuss the role of machine learning algorithms in extracting actionable insights from real-time IoT sensor data in smart factories?
Dheeraj Remella: In a fast-changing, hyperconnected industrial world, machine learning is critical. Once the most appropriate algorithm has been selected, they need to be fed the more recent data continuously to ensure an adaptive approach of the model evolution. This rapidly evolving model can then be fed into the real-time compute layer to make better decisions, decreasing false positives/negatives. These models can:
- Detect anomalies in production processes that would otherwise go unnoticed.
- Predict maintenance needs before failures occur, minimizing downtime.
- Optimize resource allocation and production schedules for peak efficiency.
The beauty of this continuous feedback loop of observe, orient, decide and act (OODA) is that these algorithms continuously learn and improve, boosting the overall efficiency of smart factories. We bring these machine-learning insights directly to bear on real-time data and decisions.
Q5. How are manufacturers balancing the need for real-time data processing with data security and privacy concerns, especially when dealing with sensitive industrial information?
Dheeraj Remella: This is a top priority. Manufacturers are employing several strategies:
- Implementing strong encryption and authentication measures. Post Quantum Encryption is becoming an increasingly interesting discussion, albeit nascent.
- Utilizing edge computing to keep sensitive data local, reducing exposure.
- Developing comprehensive data governance policies.
- Thinking of security and privacy at the initial design.
- Employee training.
- Role-based access control of the data for both personnel and systems.
Q6. What advancements in data integration technologies are enabling manufacturers to combine real-time IoT data with existing business systems for more comprehensive insights?
Dheeraj Remella: Integration is key. There is an adjacent need as well, and that is interoperability.
We’re seeing advancements like:
- Custom APIs and middleware solutions bridging legacy and modern systems.
- Standardized data formats facilitating a homogeneous target data model.
- Unified Namespace architectured de-siloing data between various IT and OT systems.
- IoT platforms are evolving to support diverse data sources and protocols, both old and modern.
- Event-driven Unified Real-time Data Platforms that make cognitive decisions powered by machine learning models on the streaming IoT data combined with near-past context.
Q7. How is real-time data processing improving predictive maintenance capabilities in smart manufacturing, and what impact is this having on reducing downtime and optimizing asset utilization?
Dheeraj Remella: The combination of real-time data and smart use of machine learning is revolutionizing maintenance. Maturation in predictive maintenance is allowing organizations to move away from scheduled maintenance, where maintenance activity is done even though the health has not deteriorated. This proactive approach is drastically reducing downtime and optimizing asset utilization across industries. This shift into need-based predictive maintenance also improves the sustainability stance and narrative at these organizations.
Q8. Can you explain the concept of “digital twins” in manufacturing and how real-time data processing is enhancing their effectiveness?
Dheeraj Remella: Usually you would find digital twins to be pure state stores that record the current state of the physical twin. This approach is quite useful in simulation and industrial metaverse-type scenarios. But, increasing maturity at manufacturing organizations is demanding digital twins to also account for the physical twins’ behavior. More often than not, there is an interest in augmenting the behavior with predictive ML models. Now, they are active participants in day-to-day operations with bi-directional data and control flow.
Real-time data processing is enhancing their effectiveness by:
- Providing up-to-the-second information on asset performance.
- Enabling simulation of different scenarios for optimization.
- Facilitating predictive maintenance and process improvements.
- Invoking actions through actuators/controllers to complete the real-time sense-control loop.
But, to make these usage patterns genuinely effective and impactful, there has to be a paradigm shift to start thinking in terms of edge computing, low latency roundtrips, data, and decision immediacy.
Q9. What strategies are being employed to ensure the scalability and reliability of real-time data processing systems as the number of IoT sensors in manufacturing environments continues to grow?
Dheeraj Remella: Increasing sensors and systems and the mounting need for intelligent automation push the narrative to the edge. While the device edge is too narrow of a context, there is a near-edge tier of computing that has the full system and process content. While there are technologies that can scale to the entire data velocity and quantity without compromising on reliability or resilience, manufacturers have to look into localized computing at the edge so that they are also addressing the timeliness and latency sensitivity of decision and response automation. Manufacturers are:
- Adopting cloud and edge computing architectures for flexible scaling.
- Implementing robust data management strategies.
- Utilizing distributed processing techniques to handle increasing data volumes.
- Data platforms that minimize latency and keep infrastructure needs manageable for environments that do not have the luxury of unlimited hardware.
Q10. How do you see the role of AI and machine learning evolving in augmenting real-time data processing for smart manufacturing by 2026, and what new capabilities might this enable?
Dheeraj Remella: By 2026, AI and machine learning will be even more deeply integrated:
- Advanced AI models will enable more autonomous decision-making in production processes with better false-positive and false-negative recognition.
- Machine learning algorithms will become more sophisticated in predictive analytics and optimization.
- Machine learning would move to the edge as well to take advantage of streaming retraining and complete the learn-predict-act cycles, all at the edge.
- We may see the emergence of self-optimizing production lines that can adapt in real time to changing conditions.
- We may even see a degree of democratization of data, decisions and automation to use natural language interaction and management of systems.
These advancements will lead to unprecedented efficiency, quality control, and responsiveness. With the increasing accuracy in manufacturing, enterprises can enjoy much better waste, power, and equipment management, thus having a better sustainability orientation.
Q11. In what applications do Volt Active Data’s customers utilize real-time data processing in the manufacturing sector?
Dheeraj Remella: Volt Active Data’s customers commonly use it for:
- Predictive Maintenance
- Real-Time Quality Control
- Adaptive Production Optimization
- Supply Chain Monitoring and Optimization
- Real-Time Asset Tracking
- Combining sensors and video data through computer vision for complete observation
Q12. Is Volt Active Data a possible solution for the challenges posed by IoT and machine learning at scale in manufacturing?
Dheeraj Remella: Volt Active Data is designed to meet 4 key requirements without compromising on availability and resiliency:
- Scale (Number of things)
- Speed (the rate at which the things generate data)
- Low latency (How quickly do you need to act on that data)
- Data-Decision Accuracy (How accurate does the correlation data need to be compared to eventually consistent systems)
Often, the decisions would need to be augmented with machine learning model inference in real-time in an event-driven manner. Volt Active Data was built from the ground up to address these requirements at any scale with predictable latency SLAs. Our customers rely on us to make sure their systems don’t miss SLAs or lose data, and also ensure that they can integrate with the appropriate downstream and upstream systems in the most efficient way.
Q13. In your opinion, what are Volt’s three most successful manufacturing use cases?
Dheeraj Remella: The top three applications for Volt Active Data are:
- Predictive Maintenance
- Real-time quality Control for early detection of defects
- Adaptive Production Optimization
There are other applications that can adopt Volt’s values as well, such as reducing the non-productive time of assets by continuous monitoring of the asset conditions, pre-emptive ordering of required spare parts before they are needed, identifying the closest technician that can run the maintenance of the asset and the closest warehouse with all the parts required for the maintenance.
These use cases require the data layer to address:
- Scale, with the ability to handle billions of events per day.
- High performance, with the ability to process hundreds of thousands of events per second.
- Low latency, where the moment of engagement is in single-digit milliseconds.
- The ability to handle complex data.
- The ability to make complex decisions on streaming data.
- Immediate consistency and accuracy.
- No data loss.
- Geographic distribution.
……………………………………………………..
Dheeraj Remella is the Chief Product Officer at Volt Active Data, responsible for technical OEM partnerships and enabling customers to take their next step in data-driven decision-making. Dheeraj has been instrumental in significant customer acquisitions and brings 30 years of experience in creating Enterprise solutions in a variety of industries. Dheeraj is a strong believer in the cross-pollination of ideas and innovation between industries and technologies. Dheeraj holds a bachelor’s degree in computer engineering from Madras University.
……………………….
Related Posts
On Use of Analytics in Products Operations at SAAS companies. Q&A with Haricharan Ramachandra
………………………..
“It’s great to get data, it can be painful to enter data. EHRs benefit all stakeholders tremendously because the information you need is in that one computer in front of you, and a lot of it is structured. The biggest challenge is getting that information in there.”
Q1. How is your long term role Emergency Physician at Brigham and Women’s Hospital helping you with your job as Chief Medical Officer and Director of Clinical Innovation at InterSystems?
Jonathan Teich: It is tremendously helpful. I get to be in the role of the clinician, and I experience so many of the trends, the stresses, the problems and gaps that our customers experience – not only clinical, but also administrative and operational issues. Having this continuous exposure to the ups and downs of the healthcare day illuminates areas where our EHR could help to improve workflow and quality. On every single shift, I see something new that could have an IT solution.
Healthcare worldwide is going through a very stressful time: burnout is increasing among clinicians, patients are having a harder time getting access to care, communication is spotty. At InterSystems we can do something about that! And the AI capabilities that InterSystems IntelliCare brings give us even more capability to address the things that I see in my work as an emergency doc.
Q2. Let’s talk about electronic health records (EHR) and healthcare information systems. What are the main current challenges in using them for a variety of stakeholders, such as clinicians, administrators, and patients?
Jonathan Teich: Simple – it’s great to get data, it can be painful to enter data. EHRs benefit all stakeholders tremendously because the information you need is in that one computer in front of you, and a lot of it is structured. The biggest challenge is getting that information in there. Especially because of regulation and payment requirements in the US and other countries, you have to enter data in different fields, different screens, different formats, often in an unnatural way that doesn’t match the flow of thought that evolves in a typical clinical encounter. There’s way too much clerical busy work that could be automated – and that work leads to stress, take-home work that eats into family time, and burnout.
There are challenges in accessing data, too, and it’s about organization. If you want to know something medically important like “what is this patient’s cardiac risk status?”, you have to go to this page for labs, that page for medications, another page for imaging and procedure results, somewhere else for unstructured insights from previous notes. As a physician, my life is a series of healthcare scenarios and “what do I do next” questions – treat pneumonia, address cardiac risk, find an available bed, balance resources, track an epidemic; for each of these, I want to have all the information I need in front of me, in one view, nicely organized.
Q3. In a recent published paper it is reported that “Many clinicians and analysts find EHRs inflexible”. What is your take on this?
Jonathan Teich: That’s another way of stating some of the challenges I mentioned.
Many EHRs are designed from the database up, not from the workflow down. When I work in the ED, I find that the EHR is very good at handling my one or two most frequent tasks – reading test results and notes, and making note entry somewhat less painful. But it’s hard to bring up information in a different way for a different purpose – show me everything about that patient’s cardiac risk, show me everything I need to make the best treatment plan for the next patient’s stroke, tell me what rehab centers are available for the patient after that. A flexible EHR would be like a good hotel concierge – understand whatever I need to know next, and give me all the right information, in one place and in a usable form.
Q4. What about Data Quality? How do manage possible bias and ensure fairness?
Jonathan Teich: In general, capturing data and displaying it back works pretty well. Data quality is a real problem when trying to do advanced things with the data – doing analytics to understand illness trends, sharing data that came from EHRs with slightly different data models, applying guidelines and rules to improve patient care.
There’s lots of partly-missing data, data items that aren’t filled in, just as a consequence of the fast-paced, often rushed nature of medical care and data entry. Or, sometimes a data element on a form is not interpreted uniformly by the many people who are filling that form. If you build an analytic model, or a clinical decision support rule, that relies on that data, it could give you misleading results and affect your decisions. I remember a time earlier in my career when the lab code for a tuberculosis test was changed, but the rules that rely on it did not change, so incorrect alerts for TB started going out; that could have had serious consequences.
Inaccuracies can lead to bias and inequity — for example, when medical centers serving a high-income group capture more data than centers serving a low-income group, or when we rely on past inequitable practice as the basis for training our future decision rules. That has happened on a number of occasions across our industry. It’s very important to think about these possibilities up front when designing analytics and decision support.
Q5. One key aspect is also Data Privacy. What are the common best practices you are aware of?
Jonathan Teich: Of course, this is a huge issue, and as you know, the solutions are partly about better systems and partly about better human practices. Fundamentally, the goal is that only the right persons see a patient’s information, see only what they need to know, and don’t spread it into other uncontrolled channels.
For the first part, role-based access control is a starting point. Smart access control, giving you easy access to patients with whom you have a defined relationship, is better; it’s a good mix of ease of use and tight privacy control.
For the need-to-know part, there are ways to restrict access to certain elements and sections of the record. For example, administrative personnel commonly cannot see certain clinical data; or some data elements may be legally protected from being used for employment or coverage decisions. On the other hand, it’s problematic to block a clinician’s access to some of a patient’s data, because any data may be important for medical decision making. So, for clinical use, often the best practice is to control at the patient level, not so much the data level.
Q6. Correlative analyses of EHR data with other external data sources can provide significant insights. What are the main challenges here?
Jonathan Teich: It boils down to making sure the data being used is correct and appropriate for the purpose. First, you need high-quality patient matching, to make sure that the EHR and the external source are in fact talking about the same person or the same group. Second, attention to data quality and interoperability is especially important here, since the EHR and the external database were populated for different reasons and may have different understanding of the same data elements.
This sort of combination also calls for extra attention to data privacy in all of the participating systems, so that the combined data doesn’t reveal an individual’s identity or expose private EHR data in unacceptable ways.
You’re right about the important insights that can be gained from such a combination of sources, if you do a good job addressing these challenges. At the population level, multi-source data can be extremely valuable for many uses – such as concentrating opioid-treatment resources in neighborhoods where there is a greater history of overdoses, or developing school lunch support programs in areas where EHR data shows malnutrition-related health problems.
Q7. You have recently launched an AI-powered electronic health record (EHR) and healthcare information system. What is it?
Jonathan Teich: What it is, is amazing. InterSystems IntelliCare is a true game-changer from the point of view of a clinician, a patient, a health quality leader, and more. It is a deep integration of a top-class AI database and engine with a full-function EHR, as well as a digital voice and image capture system – all built for each other. This integration provides true depth in every part of the record, letting you get the benefit of AI for a wide range of applications.
From providing custom summaries for many medical scenarios, to reducing the work of both structured and unstructured documentation and ordering, to streamlining communication, to orchestrating complex workflows, to mining information in text notes, we find that the AI-enhanced integrated record can improve quality and reduce burden in many ways. I said above that EHRs need to gather and present data the way that matches my medical thought process; InterSystems IntelliCare is built to address that need.
Of course, we’re continuing to advance InterSystems IntelliCare beyond the version just announced – the potential range of benefits to the care process is staggering.
Q8. Will the use of AI improve Electronic Health Records? If yes, how?
Jonathan Teich: In a nutshell, I think well-designed generative AI gives us a chance to get back to the kind of practice we once had – more focus on the patient, more time to think and talk, less busy work, better communication, less time staring at the computer screen. It also will allow us to unlock the vital information buried in clinical notes, which is so often overlooked today. It will provide a greater, easier-to-absorb awareness of what’s going on, with a single patient or with an entire population. And one of the greatest benefits will be more usable information going to the patients themselves.
Q9. There seems to be a tension between one hand delivering streamlined workflows, reduced administrative burden, enhanced patient interactions, and improved operational efficiency and on the other hand maintaining rigorous human oversight for accuracy and safety. How do you manage this?
Jonathan Teich: With a healthy dose of skepticism, and a great deal of rigor. We are very concerned about measuring, monitoring, and improving our AI systems, particularly with respect to recall (capturing and providing all of the data that it should, without omissions) and precision (not providing false, incorrect, or “hallucinated” information).
InterSystems has a large quality assurance / testing team that plays a huge role in all of our software development; they have adapted themselves for the particular needs of testing AI, and have developed test suites that we are running constantly as we continue to develop. Guided by this, we have added many guardrails in our software and in our prompt engineering to enhance accuracy and safety.
And yes, it’s always important to make sure that there is a human in the loop at strategic points. We have a variety of forcing functions that ensure that human review and final approval is obtained where needed, especially before data is entered into the record. In general, I much prefer to use AI to reduce work burden by 80% with human-in-the-loop enforced, rather than to reduce burden by 100% but have no oversight; it’s still a tremendous benefit.
Q10. Let’s focus on one of the promises of using AI in this context: Enhanced Patient Engagement. What does it mean in practice?
Jonathan Teich: As I’ve mentioned, direct benefits to patients are among the most promising areas for AI. This is a whole interview in itself, but I can summarize some important elements. As a patient in a clinical or hospital care situation, imagine that you have a constant summary of where you are in your care plan, what’s coming up next, your latest results and vital signs, the results of your CT scans – all organized in a readable and understandable fashion.
We just don’t get enough time to give you that kind of information nowadays, but it’s incredibly important. If you need to ask your practice a medical question, AI can help you compose it and help your medical team respond more quickly; or it could streamline things the next time you need to schedule an appointment. And of course, ambient AI voice dictation means that your doctor will have more time in each visit to interact with you, instead of constantly typing into the record. Besides all that, there are lots of ways that AI can interact with you directly at home, helping you to monitor and optimize your self-care, while always standing ready to send a message to your clinical team if a potential problem is arising.
Qx Anything else you wish to add?
Jonathan Teich: This is a remarkable time in healthcare. I really believe that enhancements in data handling, interoperability, and especially generative AI are going to revolutionize our healthcare lives for the better, benefitting clinicians and health workers and scientists and patients — more than any innovation that has taken place in my career. We must always be careful to protect safety, privacy, and equity; but I believe we can. I believe that the next few years could see enormous positive changes in the work of healthcare, the accessibility of healthcare, and the quality of healthcare.
……………………………………………………..
Dr. Jonathan Teich is Chief Medical Officer and Director of Clinical Innovation for EHRs at InterSystems.
He helps lead vision, design, and thought leadership for solutions to high-priority needs of providers, health systems, and governments worldwide, with particular focus on generative AI, usability and workflow, patient engagement, clinical decision support (CDS), and optimizing how EHRs can help achieve strategic clinical and operational goals.
Dr. Teich is a practicing emergency physician at Brigham and Women’s Hospital and assistant professor at Harvard. He founded the Clinical Informatics R&D department at Partners Healthcare (now Mass General Brigham), developing two generations of innovative electronic health records, computerized provider order entry, knowledge retrieval, and CDS applications, and has authored over one hundred publications and three books in the field. He serves on numerous government and industry panels concerned with CDS, healthcare quality, and workflow design; he has served on the board of directors of AMIA, HIMSS, and the eHealth Initiative. Dr. Teich also volunteers as clinical architect and designer with OpenMRS, an open-source electronic health information system serving over fifty low- and middle-income countries worldwide.
Resources
Electronic Health Records (EHR)-2025 American Medical Association.
Related Interviews
On Data Platforms. Interview with Gokhan Uluderya
On using AI and Data Analytics in Pharmaceutical Research. Interview with Bryn Roberts
………………………..
” The most successful people that I have worked with and the most successful organizations I have been a part of have been the ones that embrace learning and change as opposed to resisting it.”
Q1. You have 20+ years experience in product and engineering leadership. What are the main lessons you have learned?
Gokhan Uluderya: I think there are two very important lessons that I have learned throughout my career: first one is change is inevitable and is a constant in our lives and second one is that learning is a lifelong process. The most successful people that I have worked with and the most successful organizations I have been a part of have been the ones that embrace learning and change as opposed to resisting it. They had one thing in common and that was growth mindset: constantly learning from others, from one’s mistakes, open to change even if it went against the grain, and using these learnings to reinvent, transform, and innovate.
This is particularly important in technology sector because we are making progress at break-neck speed and we are surrounded by brilliant minds that are innovative, dedicated, and very passionate to make a difference in our lives and leave a mark in history. We are not only inventing new technology but also redefining how we live our lives, how we run our businesses, how we socialize with other people, and how we interact with the nature and the machine. Where there is constant change, there is also constant learning. This is true for organizations as well; small and large. Our organizations are constantly evolving. The organizational norms and behaviors are changing. Technological advancements are also impacting how organizations operate.
So, one of the biggest lessons for me as product and engineering leader is that we all need to continuously learn and reinvent ourselves, our organizations, and processes. Experimenting, making mistakes, failing are some of the most effective ways to learn and innovate. We should encourage these behaviors within our teams and constantly remind ourselves that very impactful things happen with small and consistent forward progress.
Q2. What is the main impact on customers’ businesses of all the transformations and disruptions that we have seen in the last 25 yrs?
Gokhan Uluderya: The pace at which disruptive transformations are happening has been accelerating exponentially in the last 25 years.
Think about it… When I started my career, mainframe-based architectures were still prominent. We went from mainframe-based architectures to SOA then SaaS, followed by the Cloud revolution. Mobile computing, IoT and edge computing brought amazing capabilities and experiences to the fingertips of every user and almost every device imaginable.
The cloud computing model enabled the builders and innovators in a fundamentally different way. The powerful compute cluster that would take 6-12 months to procure and provision in a classic enterprise environment with a very hefty price tag was now available in the cloud with a few clicks for a few hours of use for any developer. These transformations enabled big data and analytics workloads moving to the cloud and led to AI/ML being done at very large scale thanks to the scale and flexibility cloud provides. All these brought us to the age of generative AI and we are now moving fast towards artificial general intelligence.
The advancements in AI are changing the entire technology and business landscape in fundamental ways. It is changing the way we consume technology and the way we live our lives. AI-driven automation, decisioning, hyper-personalization complemented by natural-language understanding and generative AI is moving the effectiveness of technology to a whole new level.
These transformations are very exciting, and they create a lot of new opportunities. Consumers feel the impact from to shopping to dining, in communications and social interactions, in health services they receive, and in the cars they drive. Businesses feel the impact in all business functions; sales, marketing, customer service, supply chain and logistics, hiring practices are all being re-imagined with AI.
Every business is trying to adapt to and adopt the innovation coming down the pipe and fear-of-missing-out is driving them to make big decisions and investments faster than they are used to.
Q3. What are the challenges, the lessons learned, and the recommendations for both customers and tech companies that are going through these transformations and disruptions?
Gokhan Uluderya: In this kind of a rapid-changing landscape, it is hard to keep up with the change on all fronts. It is easy to make mistakes and some mistakes can be costly. Winners and losers can change very quickly. As technology vendors, we must keep in mind that our duty to our customers is not only to bring them new capabilities and innovation but also do it in a responsible, trusted way and help them along their journeys as they try to adapt and transform.
One of the key lessons learned and our recommendations to our customers is to play the long game, keep calm, and make steady progress. It is easy to go after the shiny objects and hype and fall victim to FOMO. We advise our customers to have a long-term view and strategy and evaluate the incoming disruptions carefully through that lens. It is generally a great strategy to adopt and transform in a spiral pattern: have a strong business value hypothesis, start with a small pilot, validate technology and solution, prove business value and adoption before moving to the next cycle up in the spiral.
Be agile but not random. I have seen a lot of organizations confuse agility with having no strategy, vision, or plan. Some think it is a process thing like “scrum is agile, but waterfall is not”. Agile doesn’t mean “go where the wind blows on any given day”. Agility gives the best outcomes when an organization has done enough homework to develop a long-term conviction and a plan and is able to make the right changes to their roadmap based on their experiments and learnings throughout that journey.
It also makes a huge difference to have a trusted partner that is a companion to you on your journey rather than just a vendor. Look for those trusted partners that have a vested interest in making you successful with your journey rather than vendors who are motivated mainly by selling you more.
Q4. You are the Head of Product at InterSystems. What are your responsibilities?
Gokhan Uluderya: As Head of Product for Data Platforms at InterSystems, I am responsible for leading the teams that drive our innovation in Data and AI technology space. Our Data and AI technologies power the most mission critical workloads in healthcare, finance, CPG, and supply chain. In our portfolio, we have several products that are industry-leading solutions. InterSystems IRIS and IRIS for Health lead in the multi-model, translytical data platform space. Our Health Connect product line is a leading interoperability platform solving some of the most complex integration and interoperability problems in healthcare space. We also provide several Cloud Data Servies such as IntegratedML, Cloud SQL, Cloud Document as SaaS offers. InterSystems Data Fabric Studio is one of our more recent innovations and it enables our customers to solve the enterprise Data & AI problem by building smart data & AI fabrics.
My team builds our core capabilities in this space and helps our customers be successful with our data & AI products and with their data & AI strategy.
Q5. What is your experience in working with data, AI, and analytics?
Gokhan Uluderya: Before joining InterSystems, I spent close to 5 years at Salesforce and 14 years at Microsoft at various roles related to data, AI, and analytics. At Microsoft, I was one of the founding members of Azure Machine Learning services as part of the product leadership team starting in 2015. During my tenure at Azure Machine Learning, I also received my data science certification from Harvard University to deepen my knowledge of the theory of AI and ML.
During my time at Salesforce, I was the VP of Product and GM for Marketing Cloud Personalization and VP of Product for Commerce Cloud leading hi-scale commerce services and Commerce Einstein teams. We used AI/ML to build hyper-personalized experiences for consumers and AI/ML-driven decisioning and automation for back-office applications.
Now in my role at InterSystems, I am building data & AI technologies that solve very important problems in healthcare and life sciences, finance, CPG, and supply chain management.
Q6. InterSystems has been recognized as a Leader in The Forrester Wave™: Translytical Data Platforms, Q4 2024. What is a Translytical Data Platform? And why does it matter?
Gokhan Uluderya: A translytical data platform supports transactional and analytical workloads in real-time from a single system. It also provides support for multi-model data from the same system for structured, semi-structured, unstructured data. InterSystems IRIS is a leading translytical data platform that reduces architectural complexity and provides unparalleled performance and scale for operational and analytical workloads delivering real-time processing. We currently have customers that go up to 1.6 Billion transactions per second on a single deployment in the lab environment and more than 150M transactions per second in a single deployment of IRIS in production. Many data platforms may claim translytical support by bolting multiple technologies together, but InterSystems IRIS has a unique, differentiated architecture that makes it translytical natively in one system and for that reason it provides an unparalleled price-performance value in addition to performance, scale, and reliability.
Q7. Your data platform supports multi-model natively. What does it mean and what are the main benefits?
Gokhan Uluderya: Multi-modelity is the ability to support structured, unstructured, semi-structured data models from the same system. These data models can be relational, document, columnar, vector, graph, time-series etc. There are many database engines in the marketplace today that are specialized in one or many of these data models. What we mean by supporting multi-model natively is the fact that our data platform has a unique and differentiated architecture that provides these data models from one unified common data plane. Different data models are represented in a unified object-based architecture and easily projected into different data models.
This is very significant in the world we live in today because to be able to get that customer 360, business 360, or data 360 view you need to bring many different data types together. AI and Generative AI capabilities made this trait even more important because AI can consume all of the data together to make decisions. For example, to provide AI-assisted experience for a patient, you need to have a complete view of their interactions such as the relational data in an EMR, handwritten note from a doctor, X-ray images or summary document from another visit, or voice recording from the last encounter.
Being able to manage all these data on InterSystems IRIS natively and being able to use them for decisioning is a very powerful proposition for our customers.
Q8. What plans and vision do you have for the future?
Gokhan Uluderya: The AI wave we are going through right now is a game changer. This is not because AI is a new concept or technology; it is because Generative AI and particularly ChatGPT democratized AI in a very short period. AI-driven experiences were already part of many of our experiences, however progress in natural language understanding, generative AI, and AI-driven reasoning coming together in ChatGPT suddenly brought AI conversation to our kitchen tables.
Our vision is that this democratization and proliferation of AI-driven experiences will continue at a very high speed, and it will continue to revolutionize all aspects of life and business. AI-based experiences, automation, reasoning and decisioning, “artificially intelligent” machines will become ubiquitous. I also believe this will remind everyone once again how invaluable trusted, private data is and what kind of a differentiator it is; especially in the AI-driven world.
To get the right outcome from any AI system, to make any important decision, we need to feed it with clean, comprehensive, and more importantly trusted data. As a data and AI technology company, our plan is to empower our customers to build data & AI solutions in a trusted, reliable, cost-effective way. We will give our customers the data & AI fabric that allows them to build and capitalize on their data assets. We will optimize and automate their data & AI workflows using AI. And most importantly, we will infuse AI-driven experiences into all our solutions and applications in healthcare, finance, CPG, and supply chain.
Qx. Anything else you wish to add?
Gokhan Uluderya: I would like to encourage our readers to take a look at InterSystems IntelliCare. InterSystems IntelliCare is EHR reimagined using the power of AI and GenAI. It empowers clinicians, enhances patient experiences, elevates business operations, and minimizes resource utilization using AI and GenAI.
It is a great example of our vision in action: InterSystems data and AI technologies revolutionizing our experiences in healthcare which we all deeply care about.
…………………………………………
Gokhan Uluderya , Head of Product, InterSystems Data Platforms.
Executive leader with 20+ years of product and engineering leadership experience at leading software and technology companies (Microsoft, Salesforce, and Verizon) delivering multiple 0-1 products and services that have grown to be highly-successful businesses.
Experienced in designing and developing large-scale, cloud services (SaaS, PaaS), on-premises server products, and enterprise applications driving product strategy, roadmap, and delivery managing geo-distributed, cross-disciplinary teams (product, engineering, research, customer success, design, UA, and enablement)
A technical innovator in Data, AI/ML, Personalization space and a business leader driving a business unit as general manager responsible for P&L and C-Suite relationships.
The Forrester Wave™: Translytical Data Platforms, Q4 2024
Nov 6, 2024 — This report shows how each provider measures up and helps data and technology professionals select the right one for their needs.
Related Interviews
On Applying Data Fabrics Across Industries. Q&A with Joe Lichtenberg.
On Data Fabric and Data Mesh. Q&A with Jeffrey Fried
………………………..
“Cloud is a tool, not a destination. Hybrid is the strategy that delivers both control and agility.”
Q1. What is your role and current projects at Exasol?
Mathias Golombek: As CTO of Exasol, I oversee the technical direction of our high-performance Analytics Engine, ensuring it delivers speed, scalability, and cost efficiency across on-premises, hybrid, and cloud environments. My role involves driving innovation in query performance, self-tuning optimizations, and seamless data integration to help organizations maximize the value of their analytics.
Right now, we’re focusing on:
- Enhancing hybrid data integration, making it easier for companies to run analytics across on-prem and hybrid environments without performance trade-offs.
- Optimizing our query execution engine, improving parallel processing, indexing strategies, and workload balancing to ensure consistently fast performance.
- Expanding AI/ML capabilities, enabling advanced analytics workloads directly within the database without the need for additional infrastructure.
- Improving cost efficiency, refining storage and memory management to reduce operational costs while maintaining top-tier performance.
These initiatives ensure Exasol remains the most powerful, flexible, and cost-effective analytics solution for data-driven organizations.
Q2. Choosing the right infrastructure is a strategic decision that will impact an organization for years. What practical tips can you offer in this area?
Mathias Golombek: The key to making the right infrastructure choice is understanding workload requirements, regulatory constraints, and total cost of ownership (TCO). Organizations should ask:
- Performance Needs: If real-time analytics and low-latency queries are critical, an in-memory Analytics Engine like Exasol can provide superior performance than other databases.
- Data Governance & Compliance: If strict data residency and compliance laws apply (e.g., GDPR, HIPAA), an on-premises or hybrid approach may be necessary.
- Cost Predictability: Cloud costs can spiral if not managed effectively. Organizations should model workloads and compare TCO across on-prem, hybrid, and cloud options.
- Scalability & Integration: Consider the need for seamless integration with existing tools and the ability to scale without costly re-architecture.
- Future-Proofing: The landscape evolves rapidly—opting for an infrastructure that supports flexibility (on-prem, cloud, hybrid) ensures long-term adaptability.
At Exasol, we see organizations increasingly favoring a hybrid-first approach, leveraging on-prem for mission-critical workloads while optimizing cloud usage for elasticity and burst processing.
Q3. Let’s talk about migrating to the Cloud: Why, How and What Makes Sense?
Mathias Golombek: Cloud migration is often driven by scalability, elasticity, and ease of management. However, it’s not a one-size-fits-all solution. The key considerations are:
- Why migrate? Organizations move to the cloud for agility, operational simplification, and dynamic scaling. However, performance-sensitive workloads may not see cost or speed benefits.
- How to migrate? A phased approach works best—starting with non-critical workloads, leveraging hybrid setups, and optimizing data architecture to prevent unnecessary cloud egress costs.
- What makes sense? A cloud-smart strategy rather than a cloud-first mandate. Many organizations are now repatriating workloads due to unpredictable costs and performance inefficiencies. Workloads requiring low latency, predictable costs, and high security often perform best on-prem or in a hybrid model.
Exasol supports flexible deployment, allowing organizations to run the same high-performance analytics across on-prem, hybrid, or cloud environments—giving them the ability to adjust strategies as needed.
Q4. On the contrary, what is the biggest advantage of on-premises computing?
Mathias Golombek: The biggest advantage of on-premises computing is predictability—in cost, performance, and security.
- Performance Optimization: On-prem allows full control over hardware and resource allocation, minimizing latency and delivering consistent high-speed analytics.
- Cost Efficiency at Scale: While cloud pricing is attractive for small workloads, long-term costs often escalate due to unpredictable storage, compute, and egress fees. A well-optimized on-prem solution has a lower total cost of ownership (TCO) over time.
- Data Control & Compliance: Industries like healthcare, finance, and government require stringent data sovereignty, regulatory compliance, and security—challenges that cloud providers can’t always meet.
- Minimal Vendor Lock-in: Cloud providers have proprietary ecosystems that can make data migration complex and costly. On-premises solutions allow full control over data access, storage, and portability.
For organizations running high-performance analytics on large datasets, Exasol’s in-memory Analytics Engine on-premises consistently outperforms cloud alternatives while maintaining cost predictability and compliance advantages.
Q5 According to a Barclay’s CIO survey, 83% of enterprise CIOs planned to repatriate at least some workloads in 2024. What are the reasons why companies are choosing to bring their data in-house?
Mathias Golombek: We’ve seen this shift in our own customer base. The primary drivers for workload repatriation include:
- Cost unpredictability: Cloud egress fees and unpredictable pricing models have made on-prem/hybrid more attractive for long-term analytics workloads.
- Security & control: The rise of AI and sensitive data analytics has made many organizations reconsider who controls their data and how it’s stored, processed, and accessed.
- Performance bottlenecks: Latency and performance inconsistencies in shared cloud environments make real-time analytics and high-concurrency workloads challenging.
- Regulatory compliance: Industries like banking, healthcare, and telecom face increasing data sovereignty and privacy regulations, making on-premises or hybrid solutions more viable.
- Many of these shifts and their implications have been widely discussed: hybrid strategies—where compute happens on-prem while scalability is extended via cloud—are now the preferred model.
Exasol excels in hybrid environments by providing high-speed analytics while ensuring full control over data location and processing.
Q6. According to the latest forecast by Gartner, 90% of organizations will adopt a hybrid cloud approach through 2027. What is your take on this?
Mathias Golombek: The hybrid cloud model is not just a transition phase—it’s the future of enterprise IT.
- Best of both worlds: Companies are realizing that on-prem is critical for cost efficiency, performance, and compliance, while cloud provides agility and elasticity.
- Cloud is not always cheaper: Many organizations initially moved to the cloud expecting lower costs but are now balancing workloads between cloud and on-prem to optimize spend.
- Interoperability is key: Businesses need infrastructure that integrates seamlessly across on-prem, private cloud, and public cloud without vendor lock-in.
At Exasol, we design for hybrid-first strategies—enabling organizations to scale analytics seamlessly across on-prem, hybrid, and cloud without sacrificing speed or cost efficiency.
The key takeaway? Cloud is a tool, not a destination. Hybrid is the strategy that delivers both control and agility.
Q7. In your opinion what are hybrid cloud benefits?
Mathias Golombek: A hybrid cloud strategy combines the best aspects of on-premises and cloud computing, offering organizations flexibility, performance optimization, and cost efficiency while maintaining control over security and compliance. The key benefits include:
- Optimized Workload Placement: Certain workloads—such as real-time analytics and high-concurrency queries—perform better when executed on-premises due to low-latency in-memory processing and predictable performance. Cloud resources can be leveraged for burst capacity, external data ingestion, or long-term storage.
- Cost Efficiency & Resource Utilization: High-performance engines like Exasol can minimize compute overhead in an on-prem deployment while still integrating with cloud object storage for cost-effective data retention.
- Data Sovereignty & Compliance: Many industries—healthcare, finance, public sector, and telecommunications—require strict data residency controls. Hybrid cloud enables organizations to process and store sensitive data on-prem while leveraging cloud services for non-sensitive workloads.
- Scalability & Elasticity: Organizations can dynamically scale resources by leveraging the cloud for compute-heavy tasks (such as machine learning inference) while keeping mission-critical workloads running on-prem for predictable performance.
At Exasol, we optimize for hybrid deployments, ensuring seamless data virtualization, query federation, and cross-platform analytics without performance degradation.
Q8. The most common hybrid cloud example is to use public cloud with private cloud services and on-premises infrastructure. What is your take on this?
Mathias Golombek: A hybrid cloud model that combines public cloud, private cloud, and on-premises infrastructure is increasingly becoming the standard. However, the key challenge is not just deployment but ensuring seamless workload portability and data interoperability across environments.
- Latency & Performance Considerations – High-performance analytics workloads often require low-latency query execution, which is best achieved with in-memory, on-premises infrastructure or high-performance private cloud deployments rather than public cloud services optimized for storage rather than compute.
- Data Gravity & Egress Costs – Moving data between environments introduces latency penalties and unpredictable cloud egress costs. Organizations must optimize data locality and workload placement to minimize transfer inefficiencies.
- Security & Compliance – Private cloud helps enforce data sovereignty and regulatory mandates, but integration with public cloud analytics tools often leads to security trade-offs and additional access control requirements.
- Cross-Platform Query Execution – A hybrid approach only works effectively when databases support federated query execution, virtualization, and schema bridging, ensuring that data silos are avoided, and workloads can scale efficiently across environments.
I see hybrid architectures not as a static setup but as an evolving, workload-aware strategy. Exasol’s Analytics Engine enables high-speed analytics across hybrid infrastructures by minimizing query latency, optimizing data locality, and integrating seamlessly with cloud and on-prem ecosystems—allowing organizations to maximize performance without unnecessary complexity.
Q9. What are the requirements to build and deploy Generative AI workloads?
Mathias Golombek: Deploying Generative AI (GenAI) workloads effectively requires a combination of high-performance compute, scalable storage, optimized data pipelines, and inference acceleration. The key requirements include:
High-Performance Compute Infrastructure
- Parallel Processing & MPP Architectures: Training and running large foundation models require distributed computing frameworks to optimize vectorized execution and parallel workloads.
- GPU & TPU Acceleration: Many transformer-based models rely on GPU/TPU acceleration for efficient matrix multiplications and tensor operations.
Scalable & High-Speed Storage
- Hybrid & Multi-Tiered Storage: Storing training datasets in a combination of on-prem NVMe storage (for high-speed access) and cloud object storage is a common approach.
- Data Lake Integration: Exasol’s query engine can be used to process structured and semi-structured data efficiently, ensuring high-throughput data preparation for AI pipelines.
Optimized Data Management & Feature Engineering
- Federated Data Access: GenAI models require diverse datasets—ranging from structured enterprise data to unstructured text, images, and videos. Hybrid environments must support fast ETL processes and federated queries across multiple sources.
- Vectorized Execution & Feature Store: Efficient feature engineering requires databases that support vectorized processing, indexing, and real-time transformations, with integration options for feature storage and retrieval in AI/ML workflows.
Inference Optimization & Model Deployment
- Inference Optimization & Data Access: AI workloads require efficient data retrieval and transformation pipelines. Exasol enables near real-time analytics and feature engineering for AI models while integrating with external ML platforms for model training and inference.
- Real-Time AI Integration: Using high-speed analytical databases like Exasol ensures that GenAI models can query and process real-time data without performance bottlenecks.
Security, Compliance, & Governance
- Data Sovereignty & Compliance Controls: Many AI workloads process sensitive PII data, requiring on-prem data governance while allowing cloud-based AI training.
- RBAC & Secure AI Pipelines: Implementing role-based access control (RBAC), model versioning, and explainability frameworks ensures AI transparency and compliance with industry standards.
How does this work in practice? For example, with Exasol, users can integrate with LLMs in 3 ways:
- Exasol In-database LLM Deployment:
Download your chosen language model into Exasol’s internal file system (BucketFS) and access it via User Defined Functions (UDFs). This method guarantees that your data, queries, and prompts remain securely within your environment, minimizing exposure to external networks.
- Connect to locally hosted LLM:
Integrate with LM Studio and other language model services managed within your own network and infrastructure for a balance of security and flexibility.
- API-Based Integration:
Connect directly to external language model APIs using UDFs. This option provides rapid access to the latest models without the need for local deployment, offering flexibility and speed.
We focus on accelerating AI-driven analytics by providing low-latency, high-performance query processing, ensuring efficient data preparation, real-time feature engineering, and on-premises and hybrid AI deployments.
Qx. Anything else you wish to add?
Mathias Golombek: As organizations continue to evolve their hybrid and AI-driven analytics strategies, the focus should be on:
- Workload-specific infrastructure choices rather than forcing cloud adoption where it doesn’t provide cost or performance benefits.
- Optimizing structured data processing to support AI-driven insights and decision-making while ensuring seamless integration with external unstructured data sources.
- Minimizing operational complexity by leveraging self-tuning, high-performance analytics engines that seamlessly integrate across on-prem, cloud, and hybrid environments.
At Exasol, we are committed to pushing the boundaries of analytics performance, ensuring organizations can extract real-time insights from massive datasets while optimizing cost, scalability, and security.
………………………………………….
Mathias Golombek
Mathias Golombek is the Chief Technology Officer (CTO) of Exasol. He joined the company as a software developer in 2004 after studying computer science with a heavy focus on databases, distributed systems, software development processes, and genetic algorithms. By 2005, he was responsible for the Database Optimizer team and in 2007 he became Head of Research & Development. In 2014, Mathias was appointed CTO. In this role, he is responsible for product development, product management, operations, support, and technical consulting.
…………………………………………………………..
Related Posts
On Generative AI. Q&A with Bill Franks
………………………………………………………….
On Generative AI and Consumer Marketing. Interview with Ziv Fridfertig and Madhukar Kumar.
“In many ways the Large Language Models (LLMs) have become commoditized, and the true differentiator is now the substrate of intelligence – which is data. Companies that take advantage of their data to drive richer and better experiences for their customers with AI are going to be the clear winners. “
Q1. How can generative AI boost consumer marketing?
MK: Gen AI helps marketers be more agile in fast-changing markets and do more with fewer resources. It’s about scaling your creativity, personalizing the customer experience, and optimizing campaign strategy and execution. It’s also about automating repetitive tasks – getting more of the grunt work done extremely fast. With gen AI, everything from ads to social media to email campaigns can also be thoroughly personalized and targeted. AI-guided chatbots and virtual assistants can learn from customer preferences and behavior patterns to make better recommendations and ultimately improve the customer experience. And by analyzing trends and making predictions, gen AI can play a role in audience segmentation and strategy – even modifying campaigns dynamically based on real-time feedback and data. Finally, AI tools can help marketers make data-driven decisions by analyzing data without getting an entire team of data analysts involved.
Q2. A recent McKinsey report estimates that gen AI could contribute up to $4.4 trillion in annual global productivity. What is your take on this?
MK: I believe it. Using consumer marketing as an example, the cost savings and productivity gains are real. Gen AI is already automating some of consumer marketing’s more routine and time-intensive tasks, giving creative teams more time to focus on strategy. For example, gen AI can do scalable A/B testing to test what resonates best, design concepts and layouts, or localize content for different regions and languages. In all of these ways, gen AI is a powerful tool for consumer marketers who seek to do more with less. Similar gains are happening in every industry, so you can easily see how McKinsey has reached such bullish conclusions.
Q3. What are the ways consumer companies can create value with gen AI?
MK: In many ways the Large Language Models (LLMs) have become commoditized, and the true differentiator is now the substrate of intelligence – which is data. Companies that take advantage of their data to drive richer and better experiences for their customers with AI are going to be the clear winners.
One great example of a company realizing these benefits is Skai. Skai’s AI-powered platform offers consumer and market insights and omni-channel media activation services. As with so many of our martech and adtech customers, Skai’s integration of gen AI helps them deliver more personalized, efficient, and innovative marketing solutions.
Skai was one of SingleStore’s first customers that built a multi-tenant service on SingleStore. Multi-tenant is now the most common deployment among all of our martech/adtech customers. Our partners in this deployment are AWS and also Twingo, who helped Skai deploy our solution and push the technology to the limits for optimal performance.
Q4. How does Skai use gen AI? And how does Skai use SingleStore?
ZF: We use gen AI in many of the ways Madhukar described above. It’s about eliminating the friction inherent in “walled garden media” – digital advertising and media ecosystems where the platform owner controls and restricts access to the data, content, and audience interactions within their environment. Our machine-learning algorithms and proprietary NLP break down those barriers, enabling companies to listen, predict and keep pace with the consumer journey.
We use SingleStore as our foundational database for real-time analytics and we use SingleStore Pipelines to ingest the massive amount of data that we require. SingleStore has saved our developers a lot of time while getting our users the real-time insights they need. Though we started with an on-prem deployment, we recently moved to the cloud with AWS, which has enabled us to more efficiently utilize our hardware and maximize our performance.
Q5. What are the main benefits for Skai’s clients? What are the main benefits of moving to what you call an omnichannel?
ZF: Our omnichannel advertising platform is the ultimate outcome of us listening to our client needs. We’ve been around for a while and learned that clients have their own systems and methods for marketing operations and deal differently with publishers, but there was a common ground – it’s expensive, confusing and exhausting with the fast pace of change.
Today our clients see the value of working with a single platform over multiple retailers.
It allows them much more visibility and control. The language gap is reduced. They can get insights faster, quickly identify and react to low-performing campaigns, save time with bulk operations and maximize their budget.
The platform includes smart customized widgets, real-time analysis, advanced targeting and automated actions. With this toolset, campaign managers can easily oversee their campaigns across 100 publishers. Our AI-based features inform a strategic dashboard that gives marketing directors with actionable insights on forecasting and trends.
Q6. What data challenges do you face at Skai?
ZF: Skai has quite a unique setup, with data ingestion by 30,000 Kafka pipelines across tens of clusters, and we must cater to a high concurrency of client requests, made of 80 tables joined in a single query. That tends to push database technology to its limits.
And huge amounts of fast-moving data also creates hardware challenges. Hardware is a big investment so it can be difficult to keep pace with evolving data demands and processing needs. And administering thousands of servers isn’t cheap either.
Moving this system to AWS cloud improved our infrastructure in a few matters.
With EBS performance we migrated most of our in-memory data to storage, the outcome was reduction of hardware demands by half and improving DB performance.
With EC2 Multi-AZ we built an advanced setup to endure hardware failures without downtime, and managed to make our daily backups on S3 redundant.
Q7. You have built a multi-tenant service on SingleStore. Could you please tell us a bit about it? Why did you choose a multi-tenant service?
ZF: In a multi-tenant architecture, a single instance of a software application or platform is shared among multiple customers or users, known as tenants. Each tenant’s data is isolated and remains invisible to others, but they share the same infrastructure, resources, and core application services.
With SingleStore, we managed to achieve unmatched query performance which is impossible to get with single-tenant or inefficiently too expensive. It is cost effective, easier to manage, it is more resilient, data scale becomes a non-issue and we minimized our maintenance downtimes.
Ultimately, being multi-tenant makes a bigger impact for our clients. Deployments are easier, faster, and give us more horsepower for peak or ad-hoc loads, which keeps our user interface consistently responsive.
Q8. Why did you choose SingleStore? What are the main benefits?
ZF: Speed, scale and simplicity. SingleStore had the best query performance for our platform’s main use case. In just milliseconds, it can aggregate millions of metadata rows with billions of performance events. It can handle more than 80 tables in a single query with high concurrency. We love that the data lands in near real-time from Kafka to SingleStore, it was a hard requirement at the beginning of our journey 6 years ago.
Scalability was also a key. Big clients have big data and we need to grow along with them, with amazing query performance and zero downtime.
For simplicity, I would note the single dialect. MySQL compliance means that no special treatment was needed for building queries. In SingleStore, we don’t have any issues with inconsistency in querying.
Qx. Anything else you wish to add?
ZF: On a personal note, I project a good future for SingleStore and Skai. Both companies are attentive to client needs and success and share a vision for a centralized eco-system for data processing. With this great partnership, I’m excited for the upcoming capabilities in the cloud.
__________________________________________________________________________
Madhukar Kumar, chief marketing officer, SingleStore
Madhukar is a developer turned growth marketer and an expert in product-led growth (PLG) with more than 18 years of experience leading product management and marketing teams. He has successfully implemented PLG at Nutanix, Redis and DevRev, and is a guest lecturer at Duke University on PLG and new product development.
Ziv Fridfertig, Data Infra Manager, Skai
Ziv is a Data Infrastructure Manager with extensive experience in the ad-tech industry. His work focuses on managing and optimizing complex data solutions, building scalable data pipelines, and implementing real-time streaming systems that support high-performance applications.
Core Expertise:
- Leading and mentoring teams to achieve technical and organizational goals.
- Managing diverse data ecosystems, including RDBMS, NoSQL, Big Data, and distributed systems.
- Driving successful cloud migrations to improve scalability and efficiency.
- Designing and implementing architecture for microservices and cloud-based solutions.
With a strong foundation and experience in data management, he is passionate about delivering solutions that create tangible business value and drive long-term impact.
Resources
AWS re:Invent,December 2-6, 2024: Agenda
Related Interviews
– On AI Factory and Generative AI. Interview with Ashok Reddy. ODBMS Industry Watch. March 29, 2024
– On the Future of AI. Interview with Raj Verma, ODBMS Industry Watch, January 5, 2024
………………………….
