In another conversation recording at HumanX, Ryan chats with Dataiku CEO and cofounder Florian Douetteau about the complexities of the genAI data stack and how his company is orchestrating it. They cover the early days of data science hype, the importance of the basics to enterprise genAI programs, and why a guy who coded his own garbage collector back in the day created a no-code/low-code pipeline tool.
Episode notes:
Dataiku orchestrates data stacks and lets you create analytics, models, and agents.
Find Florian on LinkedIn.
This episode was recorded at HumanX last month. Next year’s conference will be April 6-9, 2026 in San Francisco. Register today!
TRANSCRIPT
[intro music plays]
Ryan Donovan Welcome to the Stack Overflow Podcast, a place to talk all things software and technology. I'm your humble host, Ryan Donovan, and today I'm recording from the HumanX Conference. We'll be talking with the CEO and Co-founder of Dataiku, Florian Douetteau. We're going to be talking about all the data behind AI. So welcome to the program.
Florian Douetteau Thanks for having me.
RD Of course. So at the top of the show, we like to get to know our guests. How did you get into software and technology?
FD I started working in software when I was 20. I got into software and programming when I was five or six. And my first computer was in ‘85, it was an Amstrad CPC 6128. Remember those?
RD Never heard of them.
FD It's three inch floppy disk, not 3.5, and 128 was the number of kilobytes of RAM. And I started to do Basic. Software for me is indexing and such a technology.
RD And when you founded Dataiku, what was the sort of impetus to start that company?
FD We started Dataiku in 2013 with the goal of democratizing data science. This is to be put in the context where all companies can benefit from AI, from data science, but they have very little resources to go about it with. They don’t have enough emphasis to get everything done. And so our opinion was that you needed to empower people in the business, not technologies, to do more with data by themselves. And so creating Dataiku in the universe where back then it was all about Hadoop, all about Python, and we could actually enable companies with a no code type of environment to do more by themselves. These people in the business do it without having to rely too much on data scientists.
RD So you were definitely ahead of the game with the data science push. Today, everybody's data science. Back then, how big was data science in 2013?
FD It was a lot of hype in terms of people understanding that some companies were doing data science and others were like, “We're doing it, we’re doing it, we're doing analytics, but do we do enough with data?” And so there was lots of wishful thinking back then, but then it became very much concrete with companies finding ways to leverage their data and to create lots of business value, lots of tangible business value and projects. Tangible business value seems to be kind of a hype type of word, but in practice it means that indeed companies moved from just playing around with data in a data scientist type of way to actually having in production psychic learn based models that were running continuously in order to do their pricing, or creating some forecasted demand so that they could optimize their supply, or use it in order to detect fraud more automatically if they are a bank and to automate lots of manner of processes. So we've seen all of that already in the last 10 years, companies moving from a state where data was scattered and a data science type of approach where the desire to most large organizations I think are eager, already in order for predicting models and prediction.
RD Yeah. Every company seems like they have an AI data science system, but you're on the data side of it. What's the specific, besides the sort of no code access to data science? Are you gathering the data? Are you bringing it all together or what?
FD We'd be able to take kind of an abstraction layer for models or agents, and running on top of the existing infrastructure of our customers. A very opinionated approach is that we should not use Dataiku to build yet again another repository of data. We should leave data where it is. And our customer already made the choice, for instance, using one of the three public clouds. They have made the choice of using either Databricks or Snowflake, for instance, as a data platform. They would've made the choice of focusing on OpenAI or Anthropic models, for instance, in terms of LLMs, and our goal user at Dataiku is to make sure that they can actually leverage those infrastructure and any combination of those infrastructure possible to build the AI assets that they need. And so it's to be as tech agnostic and LLM agnostic as possible.
RD I've talked to some folks lately talking about just the sheer number of data sources that people have. In the orchestration layer, do you do any sort of simulated centralization or are you picking and choosing which data source you pull from?
FD With Dataiku, you typically pick and choose and can connect to your various data platform. From there, users can find the relevant data they want to use. And so what it provides in a very unique manner is not only what is called sometimes a data range, the sense of visualizing the access to data and referring of data, but it allows you to build complex processing workflows where you massage the data and you train the model and maybe you combine this data in order to build a knowledge base that will be using the RAG by an agent. We can automate all of that in Dataiku with a visual pipeline. And what's kind of fun is that this visual pipeline is fairly abstract compared to where the data actually is. Such pipeline could be underlying executing SQL queries in Snowflake or in Databricks or queries in Redshift or whatever else. And you could actually fairly easily move from one environment to the other. And I think that this ability to be agnostic, platform agnostic is very interesting for our organization. And the multiplicity of data platforms, the landscape is moving quickly, especially the LLM landscape, and so it’s a thing that is dear to us.
RD I know a lot of our audience of developers they get a little itchy when you talk about no code platforms. How do you assuage their fears that you're doing something behind the scenes that's going to mess them up later?
FD I think that when you talk about low code, it’s like when you’re talking karaoke to an opera singer. They’re like, “What? Really? That’s not real singing?” And second, there are multiple aspects. First, the people designing that type product, including myself, are, let's say, fairly much developers at heart. Meaning my first language was C++ and then I moved to C because I was like, “Oh my God, I need a real language that I can rely on.” Otherwise, I would have to pull everything on my own, including a garbage collector. Too much data. That's a mindset to reassure our audience. And it means that we built Dataiku knowing that developers will come to it at some point and have to collaborate with business people in it. And so it means that in Dataiku you've got a full IDE. You can actually run your own and choose whether you want to use, within Dataiku, Jupyter Notebook or VS Code or our studio if you want to, within the platform, all integrated within it. The projects themselves are actually through JSON files and the file, the actual files, are inside a Git repository, meaning that you actually get GIT versioning out of it. So it's a fairly developer-friendly environment, or as friendly as can be for a no code environment. And in the platform, what is sometimes very interesting is a pattern with our customers that data scientists can focus on building the piece of code that is useful to make the algorithm smart, but they can delegate the data prep to business users that actually understand the data. They can build reusable piece of code that can be combined by people in the business to get things done.
RD And the data prepping like a ETL pipeline or something?
FD Yeah, and I think that this, I think this is powerful also in the sense that lots of central IT or data science teams are sometimes overwhelmed because people in the business have requests about data that get to the central data team or the central data science team, whether right or not. And there is, especially in medium-sized to large-size enterprise, definitely some benefits from enabling more of the business to think by themselves.
RD Obviously the data is super important to the AI. I think in the press conference with Thomas Wolf of Hugging Face, he said data is going to be the differentiation as models commoditize. What do you think about that?
FD I think it would be a combination of data, but also building agents comparing uniquely data and models applying with your own business logic. In 2023, we had a year where models started to talk realistically, but where in the enterprise maybe 99% of Gen AI projects were fluff, kind of like ideas of things. 2024 models moving to fairly smart, as in surprisingly smart, and so maybe it was now like 80/20– 80% fluff, 20% real. I think in 2025 it's no longer about the quality of the models, the thing for many enterprise projects. It’s really down to managing the data, managing the what business logic to build into it. I think the next step is that for the enterprise, the big question will be [inaudible].
RD Yeah. I think that when I talk to folks in enterprise ready and they're like, “Well, hallucinations. How do we prove this out?” and I think finding that control and that reliability of something that is sort of by design unreliable and nondeterministic is going to be very hard.
FD Yeah. And also very easy to build the world. The word ‘generative’ has multiple meanings. Creating agents is not the hardest thing in the world. And if you take the perspective of an enterprise, it's very easy to imagine that each team who is in an enterprise could have 1 to 5, 10, 20 use cases. We multiply by the number of teams, business units and so forth, you can get to 100, 200, 500 use cases for agents in a given enterprise. And when you think about it, it can be very, very quickly a big mess. And a topic for the enterprise would be, how do I understand who is doing what, do we have actually the skills to maintain them in the long run? They build externally and I actually don't understand what they're doing. Do I have the proper delivery framework for that? And delivery framework means ability to maintain this in the long run, not just build, and ability to test and actually test and maintain tests for agents in the long run. So there is all of this practice from AI, which is kind of to be built.
RD Everybody obviously is talking with agents and it's interesting you talk like that. It's almost like you need these agents to be auditable. Big companies have to have audible trails. Are there frameworks, are there things being put in place for that or do those need to be built?
FD We have capabilities we're building in our platform and the ecosystem is also building a lot software tools and capabilities [inaudible]. What we believed in our platform is that there is a huge need of thinking very consciously about agent architecture. And for instance, one of the concepts we like at Dataiku is the concept of LLM image, where instead of thinking of the agent as a, let's say, blackboard application, you think about everything as in the agent and image base, LLMs, the tools, as reusable components. You think about the features, meaning the data you put on top of them also as reusable components, and you think about all of those reusable components such that they need to be logged and [inaudible] It's kind of the same concept as service based architecture that needs to be a purpose and thought in the context of where would you build agentic applications. This is key because many in-house applications need to operate in a way where the execution plan is known for AI, meaning when we look and use deep research against what we're using it to sometimes, you're using a system that potentially, in theory, power the internet ad libitum, and you start from one point and you didn't necessarily know what were the steps, and that's very interesting for some market research or creative tasks. But many enterprise processes need to actually go through a fine number of steps and value branches need also to be a finite number. Why? Because you need to be able to audit the process at the end so you actually know what the steps are, because any deviation would be another type of process that you would need to order differently. And so the way enterprises work, at least today, they need to go through this type of mindset, especially for everything in their processes that is regulated in a way.
RD Regulated is interesting because we have a bunch of finance, healthcare, AI companies where they are building off of these very important data sets, but again, these are regulated privacy. Do you think would AIs need more guardrails around their data or is it just we trust the LLMs not to leak out the data?
FD Well you need [inaudible] for various kinds. And you need the critical gatherers to protect against tools, detection and so forth. You need gatherers to ensure for content that would create liability, that could include PII content. It could include confidential information, any type of data. We need gatherers to make sure that your agents stay on track on the topic they should be focusing on, and for instance, in financial services, some agents would be allowed to get financial advice, but some others not, so we need to make sure they don't go off topic. Even gather evidence against very specifically bad data and hallucination where the LLM is pulling in facts but the fact is actually being backed by a resource. So in fact, you’ve got a specific type of governance you want to build into your application and layer in. What's interesting from an engineering standpoint is that sometimes you meet those gatherers only at the endpoint, and in many other situations, you run into some of those gatherers at every agent and the LLM that is called within the agent itself. And so it's an interesting engineering problem for companies to do that. You can try to solve it by LangChaining your way into all of this, and it works. The issue for companies is that doing that at scale across multiple business systems in the hundreds of agents will not scale.
RD And what points in the data process do you cover that? We talked to somebody where they're talking about most people process and clean and perfect their data in the ingress side, but also you need to do it on the egress side as well. Do you have an opinion on that?
FD Yeah. You need to clear your data in cycle [inaudible] quality of the output, but the question of data on the egress side for agent is, well, at the end of the day, it's related to the fact that in some situation, you don’t [inaudible] release a specific type of data or specific amount of data. If you get to a specific example, you might want an agent in a financial context that can provide some aggregated information about the market, but you don't want this agent to provide too much of a granular information because it could be information leakage even within your own company. As a regulated financial industry for instance, you might mean that some of your customer facing clients do not have access to granular information about other persons but have access to more less granular aggregate data on the market to provide good advice right to your customers. So solving for this type of problems requires some structuring of the data because you need the data to be right, but indeed you need some bit of protection to make sure that you won't go in all directions at the end of the day. [inaudible]
RD Yeah. The PII and then internal user controls is an interesting problem. For a lot of the data platforms we talked about, everybody seems to have their favorite. You need a vector database, but maybe you can have a vector database part of some other database system. Maybe Postgres is the best one. I've had people argue that. Maybe you need something, a data lake house with whatever other. Do you have an opinion on whether there is a required data platform or format for AI?
FD So at Dataiku, we are like Switzerland, very aggressively agnostic. And the reason is not because we lack opinions but because we believe that most of those opinions could be short term given how fast the market is moving. And so given a product, to give you an example, some of our customers started with Hadoop back in the day on prem. Then they moved to Cloudera, then they moved to Spark. These people didn't know the name. And then they moved all of that to Spark in the cloud, Kubernetes for whatever reason, and then they moved to either Snowflake and Databricks. So they moved across 10 or 12 years through 5 or 6 generations of the tech platform. That could be the extreme, but changing every four years seems to be normal in the industry. And so the thing that indeed there will still be a variety of data platforms in the market, and because we're for enterprise customers, we want them to be able to move to new ones and not at their own pace, and also for sometimes very basic financial reasons, it keeps opportunities when you do not rely on one platform with a lock-in where you would like the opportunity to have options [inaudible].
RD So with the Switzerland mindset, how do you consider building an orchestration platform where the data systems could change, could be new ones all the time. How do you stay agnostic from a sort of product point of view?
FD It's a lot of work in finding the right bit of attraction of what is data set, what are the data types, all the system reacts with it. Some nuances there. Oh, and visually you present, I don’t know, data prep where reduction of dates and date time be slightly different from Postgres and Snowflake and Databricks and whatever else. I’m getting old enough that I actually don't remember all of this, and I want actually to forget all of this like most people, which is where we kick in. Indeed, you want to abstract away some of this complexity from the users because it's overwhelming to get sometimes into this. So there is no magic there. It's mostly our work as a software vendor.
RD Yeah. Just sort of making all the connections and everything that's available and then abstracting that connection to be like, “Just get data. I don't care where it is.”
FD Yeah. And even to some extent the vector database and the LLM abstraction between [inaudible] simple because, well it’s text to text, text in, text out. But can have their own specificities of a parameter an LLM can take or not and so forth. So indeed, it's also to empower our customers [inaudible].
[music plays]
RD Thank you very much, ladies and gentlemen. I've been Ryan Donvan. If you liked what you heard or disliked what you heard or want to talk about data, whatever, email us at podcast@stackoverflow.com, and if you want to reach out to me directly, you can find me on LinkedIn.
FD I'm Florian Douetteau, CEO and co-founder of Dataiku, and if you want to reach me, you can find me on LinkedIn.
RD All right. Thank you very much everyone, and we'll talk to you next time.
[outro music plays]