Dataiku: The Single Platform for Everyday AI with CEO Florian Douetteau

Sure. There are many ways to talk about a company, and sometimes you can start with the numbers. And because we are a little bit about numbers at Dataiku, fine, starting with that. Dataiku is 10 years old as a company. And in terms of numbers, we are north of 1,200 employees. And we ended up last year north of $200 million of annual recurring revenue. So we sell one software platform, Dataiku, across the world and have been growing fast in the last few years.

So we started to build the platform from, let's say, the vantage point of an early CDO back 10 years ago. And I was myself able to put myself in those shoes because, back in the day, you had few enough CDOs that you could actually be hired as a part-time CDO here and there. So it was kind of like playing the mini CDO, which is a fun thing to do. And from this vantage point, the perspective was that connecting more of the domain experts, the business stakeholders, to the data science and technology teams was paramount first.

Second, the ecosystem of data was very complex, like the databases and so forth. This was a big mess in terms of the number of algorithms or things you needed to stitch together to get anything done. And third, last but not least, there was this misconception of getting multiple tools and then having to connect them together to build a full lifecycle of data to AI. And my gut sense was like, oh my God, this is very, very, very complex plumbing.

It will cost a lot. It's actually very complex to maintain and connect those things. There is lots of value if you could just have everything in one place. The easier it would be. So that was kind of like a starting point of creating the Dataiku platform and building a platform where you could not prepare your data or do machine learning, but do both at the same time. Where you could build a model, which takes time, but also, by clicking on a button, push that into production and get things up and running.

And do all of that with the mindset of being coder-friendly, as in the data scientists doing Python, but also clicker-friendly, as in someone coming from a domain that can be a finance engineer, that can be a mechanical engineer, that can be a marketing research person, and so forth, that could dig into the data by themselves with the platform. So being no-code, low-code, and being full-encompassing of this lifecycle of data to AI, which enriched a lot across the years. But essentially, it's like you get data, you transform it, you build a model, you move it to production, you build an app, you're happy.

I think we built the platform trying to find a balance between building a platform very easy to use for the business, as in someone coming from the business could click, get their data, do lots and lots and lots by themselves. The motto was being self-service from the perspective of analytics and ML. And why? Because you don't have enough data scientists actually to get things done in most organizations. But also not to be a naive platform. And especially, the world of open source was great, is great, will be great.

And so the idea was to have a platform where it was super easy to code, to run multiple coding environments, to get in your favorite IDE, whatever it is. You can choose JupyterLab or VS Code or whatever. You can all do that within Dataiku because we don't care about wars between software around us. It's all about having one place where it's very easy to integrate and orchestrate this very rich ecosystem of tools, of databases, of algorithms that is forever changing in the world of AI.

I think there are multiple definitions of the lifecycle of data and AI, but our own is very simple. It starts with Dataiku. You can actually connect various systems and databases. Actually, you can connect to most of your existing databases, whether that be cloud, non-cloud, and so forth, and you can search within them in order to find the relevant data you need for your projects. Step one, you search and find. Step two, most of the data requires some processing to get anything done.

It can be about transforming something, about parsing a date, it can be about merging two columns into one, it's about extracting some information which is inside a text field or whatsoever. It can be about correcting the data within one column because the original input is wrong and whatsoever. So that's all of these data preparation, data quality things, which can take ages. So we built within Dataiku a visual environment where you can click around to actually correct your data a little bit like you would do row by row in Excel, but actually the way you do it is by adding rules that transform your data and then can be applied to a full, very large dataset.

And here, actually, I can introduce another concept of Dataiku, which is that Dataiku is a very lazy platform. We try to do as little compute as possible. We are French at the beginning, so laziness, we know all about it. And so at Dataiku, all of this computation will be pushed down to your existing database. As in, it's not yet another system where you have to import all of the data and copy it and duplicate it. No, it's actually pushing down the compute there so that you can get refined, clean data also in your new shiny database.

Then another step in Dataiku is that you can then do AutoML, automated ML, automatic ML. I never actually got what AutoML actually was, the auto part, but essentially it's a capability of building a model automatically. And here, actually, in real life, you've got at this step a wizard because there are multiple ways to actually skin a cat or actually build a model. And specifically, you can ask whether you want to have a very simple, very easy-to-interpret-and-explain model, or have a very performant but actually slightly more black-box kind of model.

And so you say—

Yeah, and then there is no magic. The platform will actually test different kinds of models on your data, optimize performance and parameters and so forth, in order to actually get and build the best possible model of your data so that you can forecast anything coming out of your data: classification, regression, and what else. And then at this point, you've got the model. In real life, too, what you need to do is actually take a look at it to make sure that you understand the performance of the model.

What were the actual variables actually used in the model? Why is it important? It's because in our experience in real life, many people would try first and foremost to build models that make no sense from a business perspective. As in, you are trying to make a prediction out of data which does not really exist in real life when you launch to make the prediction, as in you're trying to predict the future with the future that doesn't exist, or these kinds of things.

So you either build obvious models, or sometimes you're building a model that makes no sense from a causality perspective, as in whatever thing is depending on the temperature in the room, but in fact it's correlation and not causation. So as a business person, or actually as a person, as a human being, you also need to inspect models in order to make sure that you understand that they actually make sense from a business perspective. And it's an important part of the process, both from a quality-of-the-model perspective, but also in many instances, and especially when manipulating personal data, from a bias-mitigation perspective.

So you also can do that within the platform. So I described all of those steps of things you can do with Dataiku. All of those can be done visually. But of course, if you are more like a coding person, if you want to take things into control, if you want to use or integrate your own specific algorithm or piece of code, or do some SQL for transformation, or Python, or R, or even Julia if you want to, Dataiku is actually also a very strong coding platform where you can integrate pieces of code and actually build them into very easy-to-manage projects.

And all of that leads to building within Dataiku a full data project. And a data project is a combination of several steps where you are transforming the data, you integrate them, you train a model, you use a model, you train another model, you use a model, you do some backtesting and so forth. Real data projects are actually multiple steps. And what we also provide with Dataiku is the ability to see all of those steps in one place, understanding what is doing what and who is doing what, because also in real life many data projects involve multiple people and multiple teams.

So there is lots of value in being able to segment and partition who is doing what and have this traceability, understandability, ability to just comment, discuss, and collaborate in a way that is more visual than Git, essentially speaking. Just because Git has some, let's say, limitations if you want to really have the ability to have a discussion which is linking the logic you want to apply to the data and the data. And also Git and coding environments remove any possibility to have a meaningful discussion with people coming from the business, just because they won't get into code.

So you can do all of that in Dataiku. And then comes the next part of the magic, the so-called MLOps kind of magic, as in this ability to move models to production. And here again, a misconception is that moving a model into production is about turning a model into an API, which is of course one step. But also in many instances, it's about robustifying models that can be applied in a batch fashion, robustifying the fact of automating the retraining of a model, the update of a model when the underlying data is drifting, and how you can make that in a robust, enterprise-grade manner.

And you can also have a multiplicity of virtual environments, variants of models, and actually not spin one, but hundreds of models depending on the cases and context and so forth. So, building these kinds of robust MLOps capabilities. And last but not least, we also provide in Dataiku the ability to build some visualization and apps. And here, it's not for us—the goal is not for us to build yet another version of dashboards, because I think you've got enough dashboard players on planet Earth in order to, I don't know, in order to do what, in order to fill all of our screens and our heads forever.

But it's actually to focus on the kind of apps that require this additional layer of intelligence and interaction that are linked to having predictive models. And then the kind of apps where you can input some form of parameters to play scenarios and so forth that are typically used by some more expert people in the field in order to strategize using data, making some forecasts on their sales, but also inputting their own hypotheses and running a model on top of it.

These kinds of scenarios of apps are typically not so easy to build within a business intelligence framework, and building from scratch a full web app is usually also too complex from the perspective of the enterprise. So we also help with that with our platform. So, long story of what I see being the life, regardless of the data, actually, what I see being the actual real lifecycle from data to AI in the enterprise.

Yeah, yeah. In the recent year, we started to add pre-built apps to our platform in order to further accelerate the work of our customers and help them in terms of either ideation or just delivering on those apps. But indeed, there is a variety of use cases of data and AI in the enterprise today, that's for sure. In terms of use cases, it can range from use cases in sales and marketing environments and functions, which are the usual suspects of analyzing your churn, understanding your customer profile, segmenting them, building next-best-action kind of capabilities, which are very common and that apply today both in retail environments, but also in banking environments, but also in telco environments and so forth with our platform, which are a whole niche of use cases.

But there are also more and more use cases of data and AI that are applied to more core operation-oriented manufacturing functions in recent years, where we also serve use cases that are about understanding and enabling your infrastructure or your fleet or your manufacturing operations, depending on what kind of company you are, as in, like, a large manufacturer or utility or a logistics kind of company, where you gather in one place all of your sensor data as well as your sales data, all of your operational data, and you enable people from your shop floor, people from your assembly line, and so forth to be better at building models that help understand quality control demands, the number of parts they will need, and so forth in order to meet the business demand.

And so you see this switch, actually, in the market where people having very typically a master's or an engineering degree that were using Excel a lot in order to build their day-to-day operation and build their KPIs and so forth, and were using this as a logic in order to drive the business, as in, like, how many parts should I actually order next week in order to fill the demand on my assembly line, are able now, because they've got more data and more facility to do so, to do by themselves, or sometimes with some help from a data scientist, to build an actual model that will help them get better at this kind of job.

And the beauty of data and AI is that it's a market that is becoming very mainstream, as in applicable to many industries, but where you still have lots of use cases where you indeed find very quickly strong uplift, as in, indeed, in many businesses applying models that are actually not so complicated, but still built with real-life data and with large amounts of data, you outperform your previous business rule by a significant margin.

Yeah, indeed. I think it's hard to grasp how deeply AI, machine learning, and so forth is used day to day by the enterprise. I think we get it sometimes as consumers just because we see some of this AI through some of the platforms. And actually, for people listening to us today, either through Spotify or YouTube and so forth, of course you can see the recommender system either at the right or bottom of the screen, which is some form of machine learning application.

But indeed, many business processes are including some form of machine learning, automated forecast, automated decision-making. It's been the case for a while and it's becoming mainstream. And it's becoming mainstream while not being the case for every company or every process, whereas I think that in the next 5 to 10 years, it will be actually universally mainstream, as in most companies, a very vast majority of companies will have hundreds, if not thousands, of automated systems that automate ordering and some form of cash management and some form of customer targeting and so forth in a way that is way more sophisticated than today.

As in, like, you've got AI a little bit every day in the life of the company. And today it's maybe two days a week. Like, you've got AI Monday, Tuesday, not every day.

We started Dataiku 10 years ago, which was back at the time of Hadoop. I think that most of the audience of this podcast is probably aware of Hadoop. Hadoop was an open-source distributed technology to try to store and distribute compute back in the day, still actually used and relevant today. But indeed, back then, you had lots of hope in the enterprise in order to build central data lakes, with some limitations in terms of making it happen.

But already as a company, our perspective was that the underlying data infrastructure, as in, like, where you store the data and how you compute, will actually change quite a bit, and there will be a diversity of sources of data from the perspective of an organization, even with a central data lake. And so there will be lots of benefits in being a layer sitting on top, enabling you to abstract the logic of the compute and how you manage projects and so forth from the actual location of the compute, like where the compute actually is and where the data actually is.

So already for us, it was pushing down to either traditional data warehousing systems like the Teradata or Netezza of the world, or pushing to Hadoop. And so we've seen across the years changes there, and indeed lots of adoption and movement to the cloud in terms of analytics in the recent few years, which was for us actually extraordinary in terms of acceleration for the company. And indeed, we do work very well with the modern platforms to store data in the cloud, including Snowflake and Databricks.

Our perspective, again, is to be, in a sense, lazy, as in we enable and accelerate the building of business-oriented analytical applications in the cloud. So they can actually be built faster and in a more self-service way by the business. And it helps actually create more traction and compute for the data platforms themselves.

Yeah, so we've been seeing for, let's say, quite some time a mix of structured data and unstructured data use cases in Dataiku. And that's one angle I can start with. But indeed, an unstructured kind of project would be ones involving images, involving text, and so forth. The kind of image detection, NLP kind of use cases. And it's true that typically those use cases were actually just harder from a technology perspective, requiring more data science, requiring more work in terms of data labeling and so forth compared to a structured dataset, a structured project.

And so what we have seen is that indeed companies building this kind of unstructured projects were focusing quite a bit on those where the ROI was obvious. If you're a high-tech manufacturer and you've got an assembly line building complex things, indeed leveraging images in order to automatically detect defects and changes you have to make on part of this assembly line can be very relevant because it's like a million-dollar or hundreds-of-million-dollar assembly line where winning one hour or two can actually be worth $1 million.

So building this full system of labeling images and so forth and getting into the weeds of understanding them, even if they are very specific, and building the data science around it, that's worth it. Or same if you are a big transporter and you are moving big things like planes or boats and you have to put them out of the circuit. Indeed, it's worth actually making this happen. But more casual use cases, not worth hundreds of millions, were not really a focus in terms of unstructured data.

Just because the risk-reward formula was not really working. So it was maybe only a percent, a few percent, of the use cases on the platform. I see first and foremost generative AI as an ability to actually first accelerate all of those use cases and also open up the realms of the possible, just because now with the kind of integration we've got of generative AI, many natural language tasks can be actually done faster. Because many use cases of extraction of information, including complex ones, can also be accelerated.

And also because you can imagine brand-new use cases where you use generative AI in a creative fashion in order to rethink the experience or how you deliver the data, or get one step further in terms of personalization of the output. But again, from my perspective, there is some parallel between the current situation of generative AI and the situation of data warehousing back 10 years ago, as in a multiplicity of technologies, a multiplicity of models, question marks in terms of how you integrate them together.

And I think from the perspective of the enterprise, there is lots of value in terms of abstracting this a bit. And so we built in our platform and are building integration with generative AI where it will be very easy to move from one model to the other, start with something that is using an API, and then move to something that could be more like an open-source or custom model you control and build yourself to move to production, for instance. All of those use cases where you will want to be able to build those applications at scale and have some control, just because it could be, at the end, very difficult to scale or maintain.

We do start seeing them. There is, of course, as of today, which is July '23, still lots of companies operating in discovery mode, as in wondering whether they should purchase this $3 million to $5 million contract from a consulting company to help them discover what can be the use cases or not. There is still this kind of thing happening on the market, but also there is a start of real-life usage of it, including in large enterprise.

And we see a mix of use cases that can range from end-of-the-funnel marketing use cases, where you just extend existing customer targeting applications that could be built in Dataiku, where the previous step was just to, let's say, predict the next best action for a given customer or the product to suggest, to the following step, which is to further personalize the email or the next message to send to such customer and find some uplift there, as one example.

But it's also down to lots of potential more back-office or operational functions where, for instance, in finance, in all of those areas where you've got complex document systems that are actually very valuable. Let's say, for instance, contracts in insurance, where you use generative AI, or actually LLMs, not generative AI per se, in order to extract information automatically from contracts in order to better understand the distribution of your risks and then run some analytics on it that was actually not really possible before.

But that applies to insurance contracts. That applies a lot also, I think. We see lots of demand in anything related to ESG and/or policies and so forth, where you've got lots of documentation where you want to be better at matching the text and the numbers, definitely. And then we also see lots of traction in many, let's say, research or complex environment situations where people in the field, whether it be a researcher or actually someone in an assembly line, when they've got an issue or want to explore something, are sometimes overwhelmed by information.

Quite literally, they don't want yet another dashboard just because they don't know where to look when they've got one given problem. So we've seen this ability to create a new kind of application where instead of building a fixed dashboard, you actually have a system where you ask a question, as in, "The quality on assembly line 3 is dropping. What could be the cause?" And you actually get back some information, some graphs or whatsoever that are within the context of your question, which is a new way to distribute information in the enterprise, probably more geared toward, let's say, at this current stage of technology, experts that need to get deeper into problems than the general audience in the enterprise, but which is a way actually also to distribute data and intelligence differently.

So, in all of the spectrum of an enterprise, as I say, the end-of-the-funnel sales, your back office, very finance functions and so forth, but also everything which is more like R&D or operations.

Yeah, I think that's the current situation for these enterprises, is that most of them are at this crossroads where some of their ethical values or their implications for AI are not clear. Meaning, even if you've got a strong value system as a company, how it applies in terms of AI is still TBD, and they are all hearing or waiting for regulatory frameworks that can be the EU one or the U.S. one or actually any kind of other countries that are also not fully implemented.

And so they have to actually balance a need for innovation because most CEOs I talk to agree that from a strategic perspective, they can't really wait for AI because, in all likelihood, some of their competitors would actually do better than them if they don't do it and will be just getting a competitive edge on some part of the business by being the first in terms of applying AI. So they can't wait on one side, but on the other side, they have to understand what is the risk mitigation process of AI.

And risk mitigation is about the compliance factor, but it's also about reputational risk that could get higher. It could be about just the regulatory risk of having this need to be able to report back. Then it's about understanding the control or the quality and just the business quality of especially generative AI, because potentially systems are slightly harder to control or understand. So there is this mix of all of this, and what we started to build is a framework we call RaFT to stay on top of it.

Like on top of the water, I guess, which is about thinking in terms of accountability, in terms of the fairness of the system, in terms of how transparent it is, which is essentially back to your ability to report what it does, and how reliable the system is, and to start having scorecards and frameworks to do those kinds of self-assessments, with the expectation that I think many companies will have to adopt these kinds of systems a little bit ahead of time in order to manage some of the internal expectations in terms of control of AI, or expectations from their customers, but also in order to be more ready when the compliance and regulatory frameworks will be there and not to then find some hard stop, and just have the minimum of bookkeeping and safekeeping in order to actually be able to navigate.

Yeah, well, first, applying to us, I think generative AI is also the opportunity to democratize many more things. And I see it as we apply it to our own product, this ability to add yet another layer of interaction with the user, as in talking to your computer—well, writing or asking with text—which will further reduce the cost of entry in many tasks. We see that we can help with generative AI in order to make it even simpler to transform a dataset or prepare data, for instance, or understand what the system is doing.

I think there is, in our particular case, a future of enterprise software where you've got systems of this mix of visual and text and potentially coding representation of things, where you can navigate between code and clicking and talking to your computer in terms of getting things done. And that will be a very, very powerful paradigm. That's one. Then, talking about the ecosystem in general, I think that what I see as very interesting with generative AI is that it's like a new Lego block enabling a new generation of hackers.

For many vertical applications. I think there are many, many, many vertical applications that can be built using specialized agents that combine multiple models and some business rules and just solve and try to accelerate or automate boring human tasks in a smart way. And I think there are lots of new-generation hackers that will be just doing that without necessarily having the need of a PhD in natural language processing and deep learning to do so, just a sheer willingness to solve a business problem and get into it and find the fastest way to do it.

So I think all of this new wave of innovation, there will be some limits to it and some hype and some inflated expectations, of course, but I think there will also be lots of great and interesting companies.

I'm curious if you have anything to share, whether that's your personal disposition as a human being, or if you have any lessons learned that people listening to this podcast could learn from. I think it's maybe stating the obvious, but I think as you grow the company, the easiest way not to be flying around and shooting down fires seems to be having good people you can rely upon in order to either extinguish this fire for you or not create them. And I think that when you build and grow a company, regardless of the business acumen or technical skills or whatsoever, building a team of people around you that can just keep having the same mission or motivation to solve the problem for the customer, to get things done, and so forth.

That's, I think, one of the key things. I don't know if it's a personal disposition or just the privilege of being in the field of data and AI also, because ultimately, in the field of data and AI, it was so obvious that this field is great and big and growing and so forth, that it's also very easy to keep everyone focused on the mission.

Yeah, well, first, I was never technically a data scientist. But maybe that's a detail. Definitely a software engineer or a product manager, depending on the time. But indeed, it's a big difference between being, let's say, a software engineer and being business- and sales-focused. And first, well, I'm not sure if I learned yet. Maybe still lots of things to be learned. But yeah, at the end of the day, it's a lot about making the good connections, as in finding the right people.

Ultimately, in the current world and from my vantage point, a lot of it is about learning and listening. So yeah, at the end of the day, a podcast from a great person was actually worth a lot to me in terms of putting yourself into the mindset of—meaning, many of the business topics are not difficult to grasp from an intellectual perspective. It's a lot about putting yourself into, let's say, the mindset or the discipline where it's about framing things in a specific way and sticking to it.

And the best way to understand that is sometimes listening to people that did that themselves. And then, last but not least, it's not true that you can outstretch yourself. At some point, when you come from a technical background, you won't become, like, a salesperson or a communication person day one, and maybe it's actually not an objective to become so. You have to stay a little bit what you are because ultimately you have to believe it's also a bit what people are paying for.

And it's also about finding around you people that can complement that, and actually, it's ultimately about building a team. And I do believe that building a great software engineering team or building a great product team or whatsoever is not so different from building a great team in general. Because a lot of it is about being able to understand and trust and help grow people that are potentially very different from one another.

Yeah, that's a great place to follow me and to ask for a connection.

.com. So, like, "Children of AI" with my name will pop it up on Google. And yeah, my starting point was actually, well, my kids are pretty old now. Especially my daughter is 18, which is a landmark for me, I guess. And so I realized that I was, to some extent, lucky not to have children now, as in, I'm not sure what will be the future of a kid born today in the context of AI, meaning, how do they start to learn to speak or read?

What is the impact of AI in terms of how you grow as a teenager? What are the dangers, and so forth? How do you think about your work life and career, and what are the skills you need to learn in the context of AI becoming more and more prevalent, and all of those things? So that's kind of like the mindset where I wrote this mix of an essay and fiction, imagining what would be the story of a kid starting their life today.

Thanks, Matt. Thanks again for joining us for this week's MAD Podcast. If you like the show, don't forget to give us a follow, and we'll have new episodes every Wednesday. And next week, we'll be joined by Victor Riparbelli, CEO of Synthesia, the AI video generation platform that just recently gained unicorn status after closing their Series C in June of 2023. Thanks again, and we'll see you next week.