Name: See Why Dataiku Is a 2025 Gartner® Magic Quadrant™ Leader for the Fourth Time
Uploaded: 2025-07-16T16:30:52.587Z
Duration: 1 h 16 s
Description: See Why Dataiku Is a 2025 Gartner® Magic Quadrant™ Leader for the Fourth Time

Transcript for "See Why Dataiku Is a 2025 Gartner® Magic Quadrant™ Leader for the Fourth Time": Hey there. Morning, afternoon, evening, depending on when you're calling in from. Thanks for joining us today on this webinar. While you're jumping in with your fellow attendees, feel free to introduce yourself in the chat, say where you're calling in from. I'm Christian Capdeville, and today, we're going to walk through the actual demo we shared with Gartner that contributed to our fourth consecutive year as a leader in the Magic Quadrant. So we're gonna be looking to show how Dataiku supports the end to end analytics and AI life cycle from beta prep all the way through AI agents using both code free and code first approaches. And, ultimately, this is a great way for anyone to understand how Dataiku can support your own use cases, whether that's deploying AI agents within your workflows or maybe making it easier for your business users to just make better decisions with better data and analytics. We'll touch on the full spectrum. Before we get started, some housekeeping items. So number one, this webinar will be recorded. So, will be available for on demand viewing afterwards. You'll see an email afterwards on that. And two, you know, feel free to ask questions in the q and a tab on the right of your screen. We'll answer throughout the we'll try to answer throughout the webinar. We'll also have time for q and a at the end where we can, we can dive into those more deeply. So let's get started. Maybe before we kick off with, the intro slides and things like this, it would be great to start with a question from us to you to understand some of the primary challenges that you're seeing today in your own AI initiatives. And I'll stop sharing so I can pop this poll up for everybody. And, really, this will help us to ensure that we touch on the things that are gonna be most relevant to you during the demo. We give the appropriate manner details to the appropriate kind of areas of challenge that you might be seeing today. Specifically, biggest challenges in scaling AI initiatives today might be, ML, might be agents. What kind of things are y'all running into? Appreciate the, the answers there. Looks like we're pretty, pretty spread out from, POC to deployment all the way through demonstrating ROI. Good. I mean, the good news is we're gonna touch on all these today. So we'll be able to speak to a little bit all of these and it sounds like we're, as far as the people that are here, we're seeing all these challenges. So very, very interesting. Thank you for sharing that. I'll go ahead and close that for now. And talk about what we are going to cover. So our plan for the next, you know, hour roughly is that I'll start with kind of our Gartner recognition, set the stage, kind of, what this represents for our enterprise AI, kind of offering and what should be kind of what stands out from that, which you should be aware of. I'll hand it over, to, Chad and I'll introduce everybody in a second. I'll hand it over to Chad next to explain the banking use case that we selected, and then we'll walk through the initial project setup. Then we will demonstrate the data and modeling capabilities and also the Gen AI and agent development capabilities. So we'll have at least ten minutes at the end for questions. But, again, feel free to drop your q and a in the chat anytime. So let me introduce the team that you'll be hearing from today. Myself, I'm Christian Capdeville. I'll be product marketing at Dataiku. Chad Covin is our senior technical specialist, and he led the Gartner demo project development. Erin Davis, our senior sales engineer, she will handle the data and modeling administration. Andrew Saman, our strategic data scientist, will showcase the Gen AI and agent capabilities. So each of us will kinda bring a different, unique perspective, to this, to the platform and to the story here. Alrighty. So really quick before we jump in, I'm not sure how familiar everyone is, but for those who are not super familiar with Dataiku yet, wanna give a brief intro on Dataiku. Like, who are we? So Dataiku is the universal AI platform that enables all of your people to drive value with your data. So no matter where the data lives, regardless of what the underlying tech stack is or the LLMs or GenAI services you may be using or may wanna use in the future. Those can all be unified together in our platform, all while keeping visibility and control over your data and your analytics, your AI workflows. So concretely, what does that mean? Well, essentially, we help business users create everything from visualizations to AI agents right alongside your code first users like data scientists. They work actually in the same interface to build shared understanding and ensure that your teams can ultimately drive impact with your data. And at the same time, while that work is happening in in that shared interface, we make sure IT and compliance teams are able to keep track of what gets created, how their data is being used, and enforce, you know, their unique quality, cost, and security requirements. So what we're talking about today is, the Gartner Magic Quadrant results. So we've been named the leader in the data science machine learning platforms for the fourth consecutive time. And Gartner evaluated us pretty highly on both completeness of vision and ability to execute. So we're up there on top right as as one of the leaders. We are also rated 4.8 out of five from many, many customers in Gartner Peer Insights, which to me just confirms that organizations are seeing real results and real value with Dataiku, which is our whole mission. And then the last kind of bid on this, before we move on to the meat of the of the webinar here is that there's an interesting call out here that I like to draw attention to. We're ranked number one for the product owner use case in Gartner's critical capabilities report. And to translate, what this really means in my mind is that business stakeholders can effectively drive AI value using Dataiku. So so we scored highly as well for data analysts and data scientists and model validators, which just means that all of those roles in the AI life cycle can contribute meaningfully and drive value in the unified platform within Dataiku, which again is what it's all about for us. So with that intro out of the way, I'll hand it over to Chad to walk you through the actual demo we built and to showcase some of those capabilities. Yeah. Thanks, Christian. So I'll go ahead and stop sharing that, and I'll go ahead and share my screen. And as I do, I did want to quickly remind you all of what we just saw within that poll. Right? So many people are experiencing many different sort of challenges and pains ranging from moving POC to production deployments to demonstrating, ROI to executives. And what we really wanna highlight throughout this demo is kind of a bit of each of those parts. Right? How can we actually see solutions to each of those specific challenges? And that's really the biggest part that we wanted to kind of, you know, bring forth within this Gartner demo was we built, a use case for next best offer solution for retail banking. And we really felt like this use case was able to demonstrate the full spectrum of, you know, specific enterprise requirements from complex data integrations to advanced modeling to Gen AI and agent capabilities, compliance, real time deployment, etcetera. And because banking has some of the strictest requirements in any industry, it really made it perfect for showcasing our enterprise capabilities. So let me go ahead and hide this. So right now, what you're looking at is the project home page and this is where we did much of our project management and, some of our business context understanding for, the demo. And this is also where the team got a a greater understanding. As we scroll down, we can see on the left hand side, we have a bit of, like, a summary Wiki. Right? It gives us a a good general overview of what this project is looking to do. Then on the right hand side, you can see we were actually using the, to do and the checklist to actually keep up on our end deliverables and any tasks that we needed, throughout the scope of the development of the project. Now, if we wanted to dive in, let's say we want to get a more detailed granular look at what exactly this project is looking to do, we can go to the embedded Wiki. And this is really where many of those project owners will spend a lot of their time creating a detailed Wiki so that not only their team has a good understanding of what the project is looking to accomplish, but also stakeholders have a great understanding of what is the impact, what is the deliverables that we're gonna get from this project. So, as you can see from here within this project, we are working, with the NextGen banking advisory platform. That's what we're delivering. We can see the project owner. We can see the project timeline. We can get an understanding of the summary, the architecture that was used, the processes that we used, and most importantly for stakeholders, they can understand, hey, what were the key deliverables from this initiative? As well as, if we scroll down just a little bit, it's a long document, we can also see what was the business impact. Right? What can I tangibly expect as an executive stakeholder from this initiative? So what's great about this is it's embedded right in where our team is working, so they're always keeping the end goal and the business objective in mind, but this is also great for stakeholders so that they can get full visibility on how this is impacting the business ultimately. So what's great about the wikis here as well is if we wanted to go ahead and publish this to a workspace, and a workspace essentially is just a collection of assets, and those are typically assets that you are going to share with, your end stakeholders. So, for example, you might have a dashboard within there. And if I scroll down a little bit, I can see that now that overview Wiki is now part of that end workspace, which I can then share with, end stakeholders. So before we move forward into the actual project, let's take a look at those end deliverables. We'll start with the end first in mind, and then we'll move forward, from there. So the first deliverable was an AI powered, agent, customer portal, basically. Right? It's an augmented, agent customer portal where a human representative would be within the customer portal. They would receive some sort of request from the customer. Then utilizing an AI agent, the AI agent would read the request, understand their customer information, give a proposed solution, which showed up on the right hand side, which will you'll see come back in just a moment. We'll also have an embedded Dataiku. So any questions about policy or checks, or anything of that matter, the human rep can go in and ask a question of. Then based on the agent suggestions, we can approve the changes. The agent will then suggest a follow-up, message to send to the customer after, you know, doing what they said in terms of the solution and then it will send that message to the customer. So all of this is human powered, with the augmented help of an agent. Right? So that's one deliverable. The second deliverable is a, dashboard that is all ready for executive stakeholders to understand the breakdown of the marketing campaigns for our specific target segments within our clients. And what's great about this is that we're utilizing Dataiku, which is connected directly to our trusted data sources. So everything that you're seeing within, this stories as well as this dashboard, which we're gonna hop to in just a moment, is updated in real time. So all of our decisions and all of our insights are always going to be up to date, because they're connected to our Snowflake and our Redshift and wherever else we have data stored that's trusted. So the very last thing that we have is our analytics dashboard, which just allows you to have more granularity in terms of the decisions that are made and the breakdowns of our customers. And if I go over to this customer focus, page, I can also get an understanding of just one specific customer. Right? So I can understand, hey, what's the probability that they're gonna subscribe to certain, products? What is the expected gain and expected output of that? Right? So this is great because now I can have all of these insights that help with decision making at the hands of all of my stakeholders at once in the same place where the actual work was done. So now before we hop into the actual project itself, it's important to think about, the compliance that goes into making a project like this. Right? And this answers that question of, ensuring governance and compliance. So now I'm gonna hop over to Dataiku Govern, which is essentially the governance suite that, is the overhead for all Dataiku projects. And, really, you can use this suite for three main objectives. Right? One is risk and value assessment, two is compliance and regulatory hygiene, and then three is securing and approving ops and production workflows. So first, let's start with understanding risk and value. And the way we're gonna do that is we're gonna click into, our business initiatives, sort of menu. And within this menu, we can see that we have a risk rating and a value rating, and we have a matrix for each of our, four different business, initiatives. So for example, we can look at client behavioral analytics. We see that it's a medium high risk rating with a high, value rating. So if we click into this, we can get a better understanding of this business initiative, and we can see that we have many different government projects that roll up into this main business objective. So that's what's great about governance. You can understand what projects are actually affecting what business objectives. So for the case of our example, we're gonna go ahead and look at transforming client engagement and product advisory. That's the project that we're looking to work on. And I mentioned risk assessment as well as compliance. Right? And what's really important about Dataiku and Dataiku Govern is we want to make sure that all AI that's created is responsible. And within Dataiku Govern, you can use the RAF responsible AI framework and RAF stands for reliable, accountable, fair, and transparent. Right? So before we create any sort of project or before we, do any sort of initiative, we are gonna do a pre assessment. This is what we did for our demo. We went through a pre assessment to understand, hey, what are the patterns and what are the ethics as well as the, risks of undertaking some sort of project like this. And then before we actually go ahead and create the project, it has to go through final approval. And once it's signed off on is when the actual Dataiku project can be created. So within the, actual building side of Dataiku, that's when that project would finally pop up and be created. Alright. So now as we go back to the actual governance project screen, you can see that we have and I'll actually get to this part in just a moment. You can see that we have access to the different objects that were utilized within this project so that they're always governed as well as having the proper documentation, which really helps with that compliance. So for example, if I go to source objects and I scroll down just a little bit, I can get a good understanding of what were the ML models that we utilized within our project, and I can govern them as such. I can understand what were the agents that were utilized within this project, and I can govern them as such and so on and so forth. Right? So whether it's a knowledge bank, whether it's a dataset, any object that we utilize within our Dataiku workflow can be governed, overhead. And this really helps with compliance. We can also attach any sort of documentation for each of these objects or for the whole project, within govern. And then the last part, which we just saw, is the model review. So understanding, any of the models or any of the objects that we create and before they go into ops or deployment, we can make sure that they're signed off on and we can block it from going to production if, you know, we don't give proper approval. Right? So we have to make sure that it has proper approval before it can even go, into production, within Dataiku Governance. So those are really the three ways that we utilized govern for this project. The last way is kind of going back to that risk and value assessment. We even were able to create a custom dashboard within Dataiku and embed it within our Dataiku governance so we can get a better idea of what is the projected development effort, what is the projected value and impact, and we can have it all exist with our within our governance suite. Right? So we can still get, a good view and a good understanding of the value and how it adds to the business objective. So now that's a big overview of governance and project management. So now I'm gonna go ahead and pass it over to Erin who is going to take us through the actual, data preparation and data ingestion. Thank you, Chad. Let me go ahead and share my screen. Share the paper share. Okay. Perfect. So what Chad was focusing on, which I always think is a good best practices, it showed the end first, you know, focusing on the agent powered web application and the governance of said projects. So now I'm gonna bring it back to the beginning and kind of focus on the building blocks, of a project that, you know, supplies the right objects to that web app, and to that that governed process. So what do we always start with in any project like this? It's data. Data is the gas that makes the car go. And Dataiku allows you to connect to the data you need for your use case wherever it is. So for example, we have data here, representing our revenue per customer in Azure Blob Storage. We have data representing what products each customer has in Google BigQuery. We have the data representing the amount the customer paid for each product in Databricks. And lastly, we have miscellaneous info about our customer's activities in, again, blob storage, but this time s three. And so as you can see, we're connecting to all these different, tables and all these different sources, but we know we're gonna need to be able to bring all these sources together in the most efficient way possible, which for us means being able to leverage in database compute. In this case, we're gonna use Snowflake. So we can actually use the sync recipe here, this kind of yellow bubble that you're seeing on the screen to effectively copy the data that we need, into Snowflake. So as we go forward building out the rest of our Dataiku project, we'll be able to take advantage of Snowflake's scalable compute. Now in many cases when starting a project, you may not know exactly what data is going to be right to answer the questions you're asking. So we can actually use the Dataiku data catalog to help us that with that. So here, we're able to traverse all the data that's ever been used in a Dataiku project. And, specifically, we can organize, the data that we're using across all of our projects into these curated data collections where my team and others have published datasets they've connected to or created using Dataiku that would be useful for future projects. So for example here, I have this customer's dataset highlighted within my customer three sixty, data catalog. So I can see what project it was originally used in, any tags that we have, given to this dataset, who owns it, and any other projects that it's being used in. I can preview to get an idea of whether this data is gonna be right for my use case. I can see, the results of any data quality rules that have been previously laid on top of this data and its originating project. Of course, I can also see, the schema of my dataset as well as the lineage of any column in this dataset. And, really, this is gonna be so if I look at one of these columns, let's say customer ID, This is gonna be particularly helpful when datasets that have been created via Dataiku projects are published to these data collections. We're doing some feature engineering or we're doing aggregation to create new fields. We need, as an end user to understand the origination of those fields before I go use them for myself. So you'll actually be able to see the connection of a particular column, to the recipes where it's been instantiated, or created. If I decide to use a dataset from the data catalog that's been used in another project, I'll actually see it show up in my projects with kind of a a different notation than my other datasets. They'll show up as kind of this black color so that it's clear that the dataset actually shares lineage, somewhere else with another project. Now the next step, once we've actually connected to data, what we're obviously going to need to do for any data science or, generative AI use case is clean it up a bit. But as a Dataiku end user, maybe I'm new to it, I might not feel particularly comfortable yet with my coding skills. And so Dataiku's visual recipes is a a very big part of the value that we bring is allowing someone who is a noncoder, to do coder like things through these visual recipes. In particular, I'm gonna use this very powerful, prepare recipe. It's gonna let me do just that. So I can actually point and click my way to create a script of steps, that'll going to interactively reflect what I'm doing on the data by applying them to the sample that I'm seeing here on the right as I go. Now you have over a 100 different processors you can choose, from this library. That can be somewhat overwhelming if you're just starting out with the prepared recipe. Obviously, it's great to have all those options, but you may not know where to start. So did I do those two things to help with that? We will offer suggestions on the fields that are within your sample based on what it is seeing in this column. So for example, I'm looking at a date column here. So it's giving me suggestions of processors I might use, like compute time since or extract specific date components. Additionally, you also have AI assistance embedded into Dataiku. So what you'll see going forward is how we actually leverage LLMs within the kind of batch processing of the flow itself, but we're also injecting these AI assistants to make you more productive as you're building the flow itself. So I could look at something like generate steps here, natural language, describe what it is I want the prepare recipe to do for me, and it will generate and iterate, on those processors that it creates for you. And it'll actually save the prompt that you're using here as a comment on a particular scripted step. Now let's say as I build out my project, I do wanna try my hand at using SQL to do some of this prep. Your SQL recipes will, of course, run-in database. I should point out that your prepare recipes, they will also run-in database. Your join recipes will run-in database so long as what you're doing can actually be translated to SQL. Dataiku's gonna service that to you as these little, green bubbles here. We will pick the right engine for the job, and we will even show you the query that we are generating behind the scenes as you kind of point and click your way to this script. But, again, as you get more comfortable, you may wanna use SQL recipes, themselves. Inside of these recipes, you can, of course, use the plain Jane editor that is here. But more often than not, what you'll wanna do is do some testing inside of a SQL notebook. And just like the prepare recipe, we are offering GenAI assistance to help write this SQL for me. So if I wanted to, I could, let's say, generate the prepare recipe using GenAI assistance, grab the SQL that it generates for me, push it to a SQL recipe, and then do further iterating on that SQL with my SQL assistant. The same thing can be said of, of Python code. So now that we've actually prepared our data, we're getting it ready for predictive modeling. That's when we're going to be able to actually begin the process of training a model. In this case, what we're gonna do is predict, whether or not a customer will subscribe to a particular product, with our company. Just like everything else in Dataiku, we're kinda trying to meet all the different personas that are collaborating together and coming to the table all in this one platform. So we're giving you different options for how you might, start experimenting with machine learning. The first is going to be our our visual components within the lab. So I'm doing a binary classification here. So my AutoML prediction template is gonna serve us pretty well. But I do wanna point out here you've got several other visual templates here for more specialized modeling techniques, things like time series forecasting, causal prediction or for, you know, kind of modeling, causal effects, things like unsupervised learning or either anomaly detection or customer segmentation, and then you've got some really advanced ones like object detection and image classification. So once we choose what we want to do here in the lab, say, auto ML prediction, and I'm going to choose what I actually want to model on. We can create our lab, and it will give me kind of a a templated approach to, designing here. So what I can do is experiment and iterate multiple times with multiple different algorithms, lots of different design choices here to see what's going to work best for us. So things like how I wanna do my train and test split, how I wanna handle my features. Do I want to, you know, do hyperparameter auto tuning, maybe even scaling that with something like the Kubernetes. You don't have to be a coder to do this kind of, data science work. And in fact, a lot of data scientists who are familiar with code do quite like the lab because it lets them experiment really quickly. However, if you are proficient in Python and that's the way you prefer to work, we're certainly not going to hinder you in that effort. And you can actually leverage our integration with MLflow quite deeply in order to do visual, experiment tracking on your coded up models the same way that you might be seeing them in the lab. So every experiment I run, every time I change my hyperparameters in that code rerun, we get to choose what we're logging to our experiment tracking space. And then just like we do with the Dataiku Lab, choose the experiment that's working best for us and deploy it to the flow as an MLflow model. And with both of these options, any model that you train, you're going to get out of the box interpretability and explainability plots, letting you know that how your model is performing and why it's performing the way that it is. This is especially important for, you know, business focus use cases or regulatory use cases where you're going to need to very much explain why the model is making the choices, that it is. And even more so important, if you decide to leverage these models as a tool inside of an agent, which you'll see going forward. So doing things like what if analysis, seeing the decision trees that were ensemble together for this gradient boosting model, universal feature feature importance, and then the typical kind of performance metrics that you would expect to see, in this case, from a binary classification model like confusion matrices. And all of this is publishable to Dataiku dashboards, so it's, you know, easily, experimented and and and looked at by everyone who might be contributing to this model or using it for, scoring purposes. And the last thing I wanna point out is Dataiku may not be the tool of choice for model training and model development, and that's perfectly fine too. We still want you to be able to take advantage of Dataiku from the MLOps perspective to do things like scoring, and do things like model evaluation over time. So if you are training models and saving them off as API endpoints, places like SageMaker or Databricks, you can make a connection to those endpoints and then just do your batch scoring or your batch evaluations, with Dataiku. It's also very helpful, for, your model version control. So once we actually have our model, we've trained it. We're happy with it, we've deployed it to the flow, not deploying it into production, but just making it this kind of green diamond saved object on our flow that we can always version control and roll back to. Now we can use this to, evaluate on ground truth data over time using something like the Dataiku evaluate recipe to create both metrics datasets, scored datasets, and this model eval store object. These are really useful because every time we run it, we are going to track all of these different out of the box performance metrics as well as custom one that you may wanna implement over time. So I'll be able to see how this model is performing. I can set checks on these eval stores so that if a model begins to degrade below some acceptable threshold, maybe our area under the curve has declined or accuracy has declined, this is where we would want to automate the notification to the model developer, say, hey. Maybe you need to go in here and do some different feature engineering, or maybe there's something else indicating to us why the model is starting to underperform, things like data drift. It may not be that the, model specifications have started to become incorrect, but the data that we are now scoring over time is significantly different from the data that we were training the model on. And And we can actually observe this data drift in a granular fashion to see for every feature that we've trained this model on, how is it now its distribution now comparing differently to the reference data, that we originally trained the model on. So as you can see, going all the way from, data preparation, both in a visual sense or in code using GenAI as an assistant on into model development. Again, either visually, code, or external models on into evaluating these models over time, making sure that they stay healthy and that they're going to give us the results we expect. That's when we'll be able to start to embed all these different components like the datasets, the models, and our unstructured data in documents themselves altogether inside of, one agent or maybe multiple agents and connect them through, GenAI and web apps like what Chad was showing. So with that, I'm going to pass it over to Andy. I think you might be on mute then. Yep. Hey, everyone. Hey. I'm Andy. I'm a data scientist here at Dataiku, and I helped out on building this project, specifically on the GenAI and agent front. There's a lot that you can do with LLMs and with agents, in Dataiku and in the world generally now. But we're gonna look a little bit today at some stuff we put in this project, that's new in Dataiku and might be interesting. Starting off with some of our just RAG capabilities. So for this project, we really had two, document based inputs. One was a set of documents, that look like this that really are just a hypothetical, flow of how you might do a customer interaction. So this is something that the, our LLMs and our agents are going to be able to use to kind of get some instruction on how to work with customers. And an entirely separate set of inputs is actual policy documents. Don't worry. These are all totally made up. These are not a bank's actual documents. But they represent, you know, types of documents that you might, you might have in your organization and want LLMs and agents to be able to reference, when doing work, either in web apps or in recipes on your flow. So we're going to actually work with both of them, and we'll see in a second that we can actually get both of them into an agent, in an interesting way. But just to start off, showing how you can actually get these into, what's called a knowledge bank. A knowledge bank is kind of a, representation of your information as vectors. And what that means is, it's going to be usable by an LLM. So you can pretty easily just point at a knowledge bank object, this pink square, with an LLM and say use this to help answer the question, and it will be able to do that. And when you have a knowledge bank that, is used with an LLM, you will get an object like this, this pink diamond, that is a combo of this knowledge bank. So this knowledge bank of, policy documents with, an LLM of our choosing, which we'll see in a second. So to actually make the knowledge bank, we've added and had for a while now is, embed documents recipe. What this looks like is, this really allows you to choose how you want to handle each type of document. So you can see here that different types of documents are going to be handled separately. There's a variety of different ways that you can kind of get your documents into a knowledge bank format. But one really cool one is you can use VLM extraction to kind of use a visual l l LLM, to just extract text from the document, and then embed it into a knowledge bank. You can also use any really any text extraction that you want and just, use the embed recipe to take text as a starting point as as well. Once you're in the knowledge bank, you can see that there's a lot of settings there as well. So you can choose, as always, the type of LLM or embedding model that you wanna use from the provider of your choice. So we have everything from, Cortex to, OpenAI. And you can also choose a variety of vector store types, as well. So here, we're just using face because our our data is not that big. But if we wanted to scale up, and use a much larger vector store, we can pretty easily do that as well. When it comes to actually using using the knowledge bank with an LLM. What we do is set up something like this where we would say, hey. We want to use this knowledge bank, and you can, of course, choose the LLM that you choose the LLM that you want to go with it. What we've also added is the ability to set up guardrails that go with this, rag combo of a knowledge bank and an LLM. So you can see here we've set up two guardrails, one for faithfulness, and this really just checks the kind of factual accuracy, of responses and another one for relevancy. And this kind of checks that when, our LLM uses RAG to retrieve the document, it checks that the document it's retrieving are actually relevant. If one of these checks is not passed, it will not actually execute, the reg query. And this is great because it it kind of gives you some human control over how, you know, robust you want these things to be. So if you're finding that your models are, say, not, you know, not retrieving enough relevant, information or maybe they're a little too lenient, you can kinda slide this up and down as a person and control the usage that way. So it's a really cool way that we can kinda bake governance and control, into the design process itself. All of this is, controlled by role based access under the hood. So, Dataiku is keeping track of who has access to use, use each LLM and also use this combo, knowledge bank and LLM. So you can control who has access to it on Dataiku and who doesn't. And this is pretty simple, but when actually using, when actually using an LLM in Dataiku, you have the option of, choosing the one that you wanna work with to execute in batch on your flow. A lot of times, you might want to just pick an out of the box LMM that doesn't do rag or doesn't do any of the agent stuff. But you can also choose pretty easily just to use that, retrieval augmented LMM, that we just made. This is not new. A lot of you have probably seen this. So that's pretty table stakes stuff. Something that's new, for the world really in the past year, are agents, which you can now build in Dataiku, using code, and also, visually without code. For this project, and the web app that Chad showed at the beginning, we used code based agents. So we're gonna take a look at those first. And what's cool about code based agents is that you can release whatever framework you want and also interact with all of the objects you have in the flow. So a lot of the models that Erin was just showing or the datasets that we've been looking at, are really easily able to be utilized by our agent. So if we look at, the agent that actually runs behind that web app that Chad showed, you can see it is pretty a pretty large amount of code. We put all of this in here so it's a little easier to see and we don't have to jump between screens, but you can see, you can define whatever tools you need directly in Dataiku. It can be anything from executing a SQL query on a datasets to working with images, to even structuring information. Further on down, you can see that we used lang graph. For those of you that have been, using it, might look familiar to build a pretty simple lang graph agent that kind of determines, what tools to use when. And then we were also able to define an agent, and define exactly what happens step by step when it runs, and including when it is supposed to stop and ask a human for input. We won't go line by line through this, but I just wanted to show this to kind of demonstrate what you can do if you if you're familiar with coding coding agents, and how you can do that in Dataiku. So we use LangGraph for this for this model, but this works just as well with LAMA index or CrewAI or pretty much any other code based framework. It's it's highly flexible. So this is the code heavy way of building agents. We, of course, also now have a visual way of building agents and tools, which we'll look at right now. So when building visual tools, it's pretty easy to just set them up in a few seconds. If you go here, you can see we have a few set up already. One that is a tool that runs an LM mesh query, one that runs model prediction, and a tool that does a knowledge bank search. It's easy to just add them by clicking new agent tool. You can see there are all sorts of tools. So for example, if we were to add a new knowledge bank lookup tool, You can really easily just go in and just with a few clicks choose, you know, what what knowledge banks you want to use, any additional instructions on what the agent is or is not supposed to do with it, and any other metadata that's supposed to be retrieved. When actually building a visual agent, they will live kind of alongside your code based agents so they can be governed in similar ways, and designing them is pretty straightforward. You just choose the LMM that you want to orchestrate it, and then you kind of just add tools as needed. So to add our new tool, we could just add it add it here. With all agents, it's possible to choose not just what happens, but where it runs. So right now, this is just running on the back end of Dataiku. But like any other recipe or anything else in Dataiku, you can choose to have it run-in a container if you want. And what's also been super useful and was very useful when we were developing this project, is that you can quickly test your agents, just to make sure that they are working properly. And this is awesome because you can kinda iterate quickly, but then also, pull in all the governance capabilities when you're ready to deploy. So you can see here just asking something like, can you describe the privacy policy from one of those documents, in a few sentences? It's able to do that. You can kinda see the result here. I know that this is actually pulling from, this knowledge bank on the flow, because you can actually see down here this thing called a trace. And a trace, is kind of this detailed, JSON, representation of the logic that's actually underneath, an agent call. So you can see the first thing it did is it went and determined it needed to call this tool. So it's well, it's really cool that you can see all of this trace information, see what went in and see what went out. We under understand this is probably not the best window to look at a gigantic trace. So we came up with this, really cool tool called the Trace Explorer. It's a a web app, and you can actually point it at all sorts of things. You can point it at, pretty much anywhere that you can use agents. So, whether you use them in recipes, in web apps, you'll you'll always end up with these traces pretty much as artifacts, And then you can load them into here either from a datasets or by, pasting them actually. What's cool about Trace Explorer is it takes this sort of complicated thought process for the agent and actually maps it out. So you can see that when an agent, executes, executes a query, you can see what actually happened step by step, what the inputs and outputs to each step actually were. And, this is actually plugged into the code agent that we saw. So you can see even though it's going to Langraph, which is code that I I wrote, it's not a visual Dataiku thing, it's still able to kind of keep track of, what tools were used and and what happened under the hood. We didn't use any multi agent approaches in our project, but, it's it's pretty easy to just kind of build that in, into this approach. So just here, if you wanted to, say, have an agent that queries another agent, you can have this LLM mesh lookup tool that queries another agent and then quite easily add it as a add it as a tool to a visual or code agents. Can pretty easily add it, and then kind of describe when it should be used. So it's pretty easy to just build multi agent frameworks right away, and the traces and the governance and the scalability will all kind of carry on. Another way that we introduced to be able to use multi agent approaches is, a very cool new web app called Agent Connect, freely available. If you've used answers before, this interface probably looks very, very similar. It's a chat app. I'm sure we've all seen these before. And you ask a question and it leverages some GenAI magic under the hood. The reason that this one is a little bit different than a lot of others is that when you're setting it up, you can actually choose a bunch of different projects and a bunch of different agents, for it to connect to. So here we have two different agents, and you can actually add descriptions for both of them, about when each one is supposed to be used. And then your users can actually come in and have one interface or they can ask a question, and agent connect will use those descriptions to kinda determine what is the best agents to answer the question. So you can really get a lot of knowledge across your entire organization, in one spot for people to work with very quickly. And, of course, all of this, really cool. You can choose where all the conversations get stored. So we're storing it in this, conversation results. And that means that whenever anyone uses your chat apps, you actually get them as a dataset on your flow, and you can run all sorts of things with them. So you can, you know, run the traces web app, but you can also run this GenAI evaluate recipe. What's awesome about the GenAI evaluate recipe, this one with little star on it, is it will, take your inputs and outputs and kind of evaluate how well your LLM is doing. So again, whether you build the input dataset from a recipe on the flow or from, you know, people just asking questions naturally on a web app that you have out there in the world and it's accumulating, you know, this conversation history. You can just kinda point point it at this evaluate recipe, and it will compute, metrics for how well your LLM is doing. So this is stuff that I think would be pretty hard to code out, at least for me, on my own power to do this. These are all pretty pretty new capabilities for the world. But you can kinda do them visually here. And you can also design your own evaluation frameworks with code, if you want. And the outputs will be a regular eval store in Dataiku. All of this is also stored as a dataset representation as well, optionally. But you can keep track of these metrics and kinda see how they're doing in your apps over time. So, again, a really nice way to just put a lot of things out there in the world either as visual recipes, in answers, in agent connect, in custom web apps, but you'll always get the traces, you'll always get the guardrails, and you'll always get the ability to evaluate it afterwards. So that's a little bit of what we did for this project, with agents. And I think next, I am passing it to, I think, Erin to talk about, a little bit about how we got this into production. I think we're gonna be passing it off to Chad, if I'm not mistaken. Oops. Oh. To wrap out of that. To chat to Chad. No worries. Yeah. Not not sure we have enough time to to walk through all the nitty gritty of production, but, that's essentially what we that's essentially what we, I'm hearing a a big echo. A big echo. I think it's just you, Chad. We're good. Okay. Alright. Well, Christian, I'll go ahead and pass it over to you so that you can, kinda walk us through some of the last couple of things the last couple of things. Absolutely. We can we can wrap it up. We wanna leave just a couple minutes for questions as well. I think that it's good. Now that we've kind of, covered all all of the the demo itself, I thought it would be useful to just recap, like, what did Gartner say about us? What are the important things to take away here? And, ultimately, it comes down to kind of these three strengths that Gartner highlighted for us. And the t the TLDR here is that the strengths that Gartner called out really, to me, reflect our differentiated customer centric approach. So we're bringing innovation to customers, but we're trying not to lose focus on their core use cases. So, if I were to go through these one through three, the first one is really saying that, our our focus on core data science is a strength, meaning that Dataiku is continues to be strong on core data science capabilities, including data insights, but we're also delivering future focused capabilities through generic applications and AI agents. Number two, this one's interesting because we were the only vendor to be called out for this. Dataiku's customer support and services essentially help get customers to value quickly. So I think this says a lot about our focus on making customers successful. We were the only vendor called out for this strength. And number three, counters being ahead of the market. So this is shown by, for instance, by Dataiku providing the LLM mesh and GenAI governance and model optionality that we talked about really ahead of peers. We we launched that last year because we knew that enterprises were gonna need those things to scale and operationalize, AI. So if you wanna dive deeper on, like, these strengths and some of the other things that are in the Gartner, report, it could be useful maybe to hit this QR code. If you haven't already done it, you can go and download the Gartner Magic Quadrant from the Dataiku site. And I'll leave this up kind of as we transition into, to a few q and a questions in the last few minutes here. So appreciate it, team, for the great overview up till now. And maybe I will start with one question that I saw come through that's probably very useful for people all the way from analytics to AI agents is that, you know, can I tag certain data items as, like, crown jewels to help with things like data governance and and tracking usage? And we answered this in the q and a, but, Chad, maybe you wanna speak to that question for the rest of the group here? Yeah. Definitely. So the first thing that I would say is, definitely, yes. You can tag, certain datum data items, but, really, you can tag essentially everything in data. I could right? Like, whatever asset, whatever item you have, you can tag it so that you can easily find it. You can also put your, I think it's described as what, crown jewels, those crown jewel golden datasets. You can have them within the feature store, which is natively built into Dataiku, which then you can also tag within the feature store as well. But you can also add it to the data catalog. And within the data catalog, you can create, certain data collections. So for example, one of those data collections might be all of your crown jewel datasets that you know you're going to go back to and reuse, often. So, the answer to that is definitely, and there's many ways to do it. Awesome. Another question, that I thought might be useful to raise is the control of access. So the control controlling who can access what data in the l l through the LMs. Is it via single sign on or or how does that work? And maybe, Chad, I'll ask you to lean in on that one again. Yeah. So I can start off and and then, Erin, I can pass it over to you as well. So in terms of accessing, LLMs, it's pretty similar to how you access just data generally. They're kind of treated, fairly similar within Dataiku. You know, you have a connection to a data warehouse and a connection to an LLM. So how we often do that is we utilize, what we call credentials. So you can either give global credentials or you can give per user credentials and included in that, can be a single sign on where the user has to then go to let's say, for example, we're utilizing an OpenAI connection. They have to go into OpenAI, put in their, workplace, you know, email and and do their single sign on process for OpenAI, And then that can give them then access to their LLM connection on the Dataiku side. I know we have something similar internally set up for, you know, Snowflake and for Databricks and for different connections like that as well. So it's fairly similar. And, Erin, I'll pass it off to you as well. Yeah. So, in the same place that you're creating connections to data, it's the same place that you create connections to your LOM service providers. Typically, it's gonna be via API token authentication. Like Chad said, that can either be the global token that you're gonna allow certain groups to use. So you'll define data groups that can use this connection, Or what you might have is a per user connection where there is no global API token in each individual user that's been given access to that connection. They still have to be a part of the right Dataiku group, that's been given access to that connection. They will provide their own API token in their user settings, and they'll do that little dance. For data sources, that's, more likely the place that you're gonna do the SSO dance. So if we take Snowflake as an example, you might be using OAuth two point o. We set it as per user. The end user would go to their end user settings. They would click through to connect to Snowflake, do the whole SSO dance, click on their email. It authenticates them, and then they'll have to refresh that every so often. I think someone asked a question about, like, using an LLM on data. You're gonna have to have access to both the LLM as a service provider, and you're going to have to have access to the data source itself, to bring those two components together. So, hopefully, that's clear. Awesome. Appreciate it. I thought maybe it would also be interesting to lean on maybe at least one more question here around, you know, when we talk about LMSH, how should we think about that versus something like MCP? And I don't know. It's either Erin or Chad. Which one of you wanna take that? MCP is a relatively new topic for me. My understanding of it is this is kind of a standardized way to communicate with the tools that you maybe have developed elsewhere, and you kinda house them in this m p m MCP server. I know that Dataiku, we're creating plug ins in order to connect to existing MCP servers, as well as develop your own from within Dataiku. So you if you've got external tools, you can use them. The specific question around, kinda how it's different is, it's not necessarily that we're giving that, like, totally different methodologies. It's more that we're kind of wrapping up that communication in governance and giving that extensibility to all different types of skill sets whether you're a coder or not. But I don't know if if Christian, Chad, or or Andy have anything different to say there. Yeah. Maybe the one thing, that I would add is that, on a simple level, like a high level, MCP is typically you're talking about that when you're talking about connecting agents to datasets and different systems. And when we're talking about LMS, it's really about connecting to the LLMs and GenAI services themselves to make those available in a controlled way to your builders. And that's kind of like a different goal. And LMS does that in a way that is basically making GenAI and and AI capabilities, usable in a compliant way to to the builders who are gonna be creating, let's say, agents that might then use MCP servers to access data. So it's part of the system, but it's kind of a different purpose. Awesome. I know that, I think we've answered most of the questions in the chat and, spoke to a few of them. So I'm glad we had time for that. I think that we are at time, so I might go ahead and wrap it up. Thanks again to our wonderful panel speakers here for walking us through that awesome demo. Again, there will be a wrap up email that goes out with the recording for this, webinar, so you can check out these details afterwards. Again, go ahead and grab your copy of the Magic Quadrant report if you haven't already, using this nice little, QR code right here. And we will see you all on the next webinar. Appreciate it. Thanks, everybody.