Live Demo #2: AI-Powered Investment Research in 30 Seconds Speaker: Laurent Fabre, Field CTO — South EMEA, Databricks Event: Paris — Market Data x AI (Finteda / FactSet), March 12, 2026

---

Laurent Fabre: I'm the Field CTO for South EMEA geography and I'll be presenting today what we've been doing and cooking with FactSet. I hope you know it. My background is in engineering. I worked as a kernel developer in C/C++ and more especially cryptography. And I worked in data most of my life, actually. I very recently joined Databricks, one year ago. I'm actually covering for two of my colleagues, Roman Ostrowski and Antoine, that some of you might have met before across the globe. So if I make a mistake, it's on them.

So how many of you do know Databricks? Can you raise your hands, please? That's a lot of people. Thank you so much.

Our company has evolved quite a bit over the last five years. You might have known us for the managed Spark ecosystem and the fact that we created Spark 13 years ago. Our co-founders are still teaching at the moment in labs across the world, mostly Berkeley and Stanford. And that comes with a special kind of prize that I'm going to show, at least part of it today.

Some of you might have seen that we just released a research paper. I'll talk about it very briefly because some of you are interested in research. The name is CARL. It's not a special thing, to be honest. It's something that we're going to disseminate on the platform, the Databricks platform. It's a special technique that we use to use small models, because we extensively talk about the latency, the cost of LLMs, the hallucinations and so on. So one of the key components to reduce and make it usable and cost-efficient is to drastically reduce the number of steps, the amount of reasoning, the latency, the number of tokens and so forth. I can send you the link if you're interested, or you can look it up on Google. Of course, at the end of the presentation you get my contacts and you can always ping me on LinkedIn or by email. I'll be very happy to.

So, back to the demo. Imagine that you're an analyst. You have a couple of hours, three hours to decide how your portfolio is going to be rebalanced because something came up. You're facing tons and tons of data. We all did at some point. We're talking about silos, we're talking about different indicators. Some are high confidence, some are low confidence, some are structured data, tabular data that you might have ingested through CSVs and Parquet and so on. And some of them are just plain old PDFs, mini briefs, news that you run daily through your systems, on your notebooks or just simply on some Spark servers somewhere to run sentiment analysis or any kind of attribute extraction from those data — alternative data, I might say.

So let's say for the sake of this demo, my avatar is going to be Pierre Chartier. Some of you may know that particular person. So I'm going to connect because my enterprise directory is connected to the Databricks apps, and immediately what you're going to see is a trading desk. I'm not going to teach you guys. Looks like the novelty is that it's entirely connected to the Databricks ecosystem. So you don't have middle software, middleware. It's running straight on the platform. So there's no problem for communication.

One of the key components that I'm going to show to you, assuming I can aim for the right — there you go. So that's part of the backend that my colleagues have cooked. It's based on FactSet. So we have a lot of types. Indeed, it's called Unity Catalog. Some of the distinguished colleagues I've mentioned catalogs before and how important they can be to get quality out of your data and valuable decisions, actionable decisions. So that's part of the catalog and it combines all sorts of things — SEC filings, earnings calls, news, everything you can possibly think of. And so we integrated everything. I think the date is somewhere in February. So that's not live, that's a sample that we got from our friends at FactSet.

And so we combine all that in what's called the Lakehouse architecture. You might have heard of data warehouses before. It's something that is going to fully enable the time traveling, the ACID transactions, the file tech — you can travel the data very efficiently, but only very limited to tabular data. So it's absolutely not tailored for unstructured and alternative data. What we do at Databricks is we combine those tabular formats and tabular techniques and we use a different technique to speed up the process of integrating and indexing and extracting information, valuable information from both structured, semi-structured and unstructured data. So that's what you're looking at at the moment. Granted, it's not very visually appealing, but it's just an introduction.

So what you're seeing there is just a mix of different techniques and it's live. When I'm going to click here, it's going to run an analysis of my exposure in my portfolio. And it's not, contrary to what most people think — the platforms nowadays are not limited to just visualizing, ingesting data, pivoting, wizarding and so on. You can actually act, you can actually engage your sales, your traders, everyone, every function in the enterprise. They can embrace the platform very easily. And it's multi-turn, as you can see. It's not your typical chatbot. It's what's called an MCP app. It's an extension of the MCP protocol. And so even though we use the FactSet data, we don't use yet the MCP from FactSet — it's all available on the platform in the marketplace, I believe. And so you could, of course, deploy. I'll be very happy to send the demo to whoever is interested. Of course it's free, so make the most out of it. And so you could always connect the MCP server to make it live. Actually that would be a really good thing to do.

And so here I'm going to approve. Wait, can I — yeah. So this is going to effectively modify the data and the tables in my portfolio.

There's a variety of use cases. But what I really like to do is show you why and how we did that. The pain, of course, is the sheer volume of data. Ever-changing data, ever-changing algorithms. Every day something new — new LLM, new technique, a new way to look at tabular data and get the most out of it.

There are different types of pain that you're all very familiar with. We mentioned that before. So I think the most important part is the architecture. Here's a brief outlook on what we used to cook this demo. That's a very short sample, of course — only 20 companies, a bunch of documents. It's big enough for anyone to face issues, whether it's on your laptop or data platform. It's just big enough for those problems to pop.

So this is how we get from research to execution in 90 seconds. We just demoed this part. Again, we don't necessarily do that kind of demo, customized demo, every single time. The reason why I'm doing it is because techniques are evolving and we are evolving with them. Nowadays at Databricks, we figure that we have to adopt a builder's mentality. We cannot only showcase very static data and dashboards. That belongs to the past in our opinion. So what we do is we in fact create demos, specific demos such as this one, using the data from our partners and customers. And I did this particular one in roughly two hours. If any of you is interested on just how we managed to do that, I'll be very happy to discuss, of course.

So everything that you've seen — keep the sanity of whoever is in charge controlling that — meaning we always keep a human in the loop. That's for now at least something that is prevalent in our industry.

And this is how, ladies and gentlemen, we transform the goals in FactSet into standard and into action.

Just a simple word on Databricks. I don't necessarily like vanity metrics, but I have to. So at the moment, we did some fundraising and we are very lucky to be able to invest 30% of our earnings. At the moment we have a $5 billion run rate going up 70%, I believe. And we use most of this money not in marketing, not in any way for employee morale or whatever. We invest every single dime in research. So that's why we might be able to do things fast and disrupt and innovate in this field.

You might have known us, of course, for different software that we created, whether it's MLflow, Spark, Delta Lake, Unity Catalog, but the Databricks platform brings every single component that you might have seen, especially the open-source versions. It adds a top layer of security and control and governance, which is extremely important in our industry.

And that's it. Do you have any questions?

---

Q&A

Audience Member: Have you included any memory in the execution and recommendation agents?

Laurent Fabre: Could you restate your question? I think I got what you meant, but just to make sure.

Audience Member: Have you included any memory in the execution and recommendation agents?

Laurent Fabre: No, we did not use any. We use a combination of vector search, especially for unstructured data, and attribute extraction and similarity, of course, but we don't stop there. That's like V1 for everyone — retrieval augmented generation and so on. So we don't stop there. That's where the proof of concept goes to die, in my opinion.

You might add on top of that what we call Genie. I can actually show you. Wait, I think it's better if I do — there you go. So Genie is effectively an agent of a particular type. It's called a composed AI. It's not just an LLM. It's going to look at your tables of data that you can see on the right. See, here you're going to find the tables that I was showing earlier. So the app effectively is mostly calling this Genie space, as we call it. And Genie is a special type of agent that is able to look at your schemas, relationships, the snippets of SQL and Spark code and figure things out. It's going to use AI to document your schema. It's going to use AI to find the right synonyms, sample your data — all of that automatically and at data scale. So it's a nice piece of technology.

What I did in this demo is that I managed to combine this part with the retrieval augmented generation, namely the documents and the vector search. And on top of that I used supervising agents — this very model that I was referring to in the research paper. So we use very efficient supervision, which means that the cost is not going to explode, because managing budget I think is one of the key aspects. As soon as you solve hallucinations and that type of problem — prompt injection and all the other really painful problems around LLMs and alternative data — I think you're left with budget, pretty much. So we're working very efficiently on that at the moment. We are making solid progress at the research lab in Berkeley. And this is what I showed today — we divide by five or six the price of every request. So I think it's substantial.

Audience Member: And how about the latency?

Laurent Fabre: Well, that's a very good question. Thank you for asking. When we divide the cost, we divide the latency — it's just that simple. So it's just a lot faster, as you could see. Any other question?

Audience Member: I didn't read the paper yet, but I saw in the abstract you claim you propose a new post-training paradigm. So last year there was work from Thinking Machine Labs — all the quality generation stuff. Do you use similar stuff? I want to understand your model. The backbone is still transformer, right?

Laurent Fabre: Yeah, yeah.

Audience Member: And then for the post-training stuff, did you do anything different than Thinking Machine Labs? Because for them, we know reinforcement learning is too expensive for online search and trace. So there are issues. Did you do something different for your paper?

Laurent Fabre: We did a lot of things differently, but the paper was released two days ago. I confess I haven't had a solid chance to have a good look at it. So I'll be very happy to discuss — we can exchange contacts and I can hook you up — but I'm not able to speak to that in detail. It would be too premature. Thank you for understanding.

Audience Member: How is CARL comparable with the agentic approaches like the ones from OpenAI or from Anthropic?

Laurent Fabre: So essentially it's a similar technology, as we were discussing earlier. The main difference is that, for instance, OpenAI and Anthropic tend to make a lot of steps when they reason, especially on office work. That could mean a lot of things including looking at PDFs or tabular data, it doesn't matter. We came up with a benchmark called QA Office, which is absolutely not flattering to anyone, including us. And that's the point — we are not looking for flattering benchmarks. That's not what we're after. And so what we did is use smaller models, cheaper models, and we used a special technique of reinforcement learning that we were discussing to train and make sure that we divide the number of steps on average by six. So that's a lot of latency that you get back, and essentially it transfers automatically to cost efficiency. There's a lot to unpack here, so I'll be very happy to discuss, but I think it explodes the scope. You can take it offline.

Audience Member: Yeah, sure.

Laurent Fabre: I'll be very happy to. Okay. Merci beaucoup. Merci. Perfect timing.