Video: From Business Intent to Production AI: Introducing Dataiku Cobuild | Duration: 3516s | Summary: From Business Intent to Production AI: Introducing Dataiku Cobuild | Chapters: Welcome and Introduction (14.815s), Introducing Cobuild (185.77s), GenAI Coding Challenges (253.895s), Cobuild Overview (391.76s), Cobuild Value Areas (544.81s), Cobuild Capabilities Overview (680.085s), Setup and Activation (833.545s), Cobuild Use Cases (962.66s), Dataiku Platform Overview (1078.105s), Model Training Process (1219.989s), Model Performance Analysis (1387.64s), Data Quality Checks (1522.285s), Automating Retraining and Inference (1620.305s), CoBuild Customization Options (1787.46s), Workflow Refactoring (2020.605s), AI Documentation Generation (2318.42s), CoBuild Best Practices (2493.35s), Q&A Session (2781.84s), API Deployment Q&A (2996.475s), Language and Configuration (3149.555s), Cobuild Security Permissions (3271.385s), Closing and Conference (3387.89s)
Transcript for "From Business Intent to Production AI: Introducing Dataiku Cobuild": Hi, and welcome. Thanks for joining today's webinar on CoBuild. Just a short message to let you know that we'll get started in a few minutes. So again, if you're just joining, welcome to this COBild webinar. We'll get started in a few minutes. We're just gonna wait for more folks to join. As we're waiting, go ahead and take a moment to say hi in the chat if you'd like and tell us where you're joining from. I will go first. My name is Ashley, and I'm joining from Paris, France. A couple of housekeeping items. We will send you a copy of the deck that we will be presenting today as well as a recording in an email tomorrow. So if you, missed out or if you have some coworkers who missed out on the session or you'd like to review some of the material that we shared, you'll be able to do that tomorrow. We'll wait a few more minutes, and then we'll, we'll get things kicked off. Alright. Well, it's great to see that we have people joining in from, all over all over the planet. Always great to see people, joining from all the various countries in which we have Dataiku users. I'll go ahead and get started, and, let's dive into this Cobuild webinar. My colleague, Pat, and I have the great pleasure of being here with you today to introduce Cobuild, Dataiku's new agentic assistant. My name is Ashley Wiseman. I'm on Dataiku's product team. I worked on Cobuild, and I'm here with Pat, who is a solutions architect within our platform strategy team, and we're gonna talk about Cobuild with you today. We'll keep it simple, and we'll start with an introduction to Cobuild, the value that it can bring to you and your team, what it does, how you set it up. I'll cover this first section. Then Pat is going to walk you through a demo so that you can see these concepts applied directly, and you can see Cobuild in action if you're not already using Cobuild today. We'll leave some time for questions at the end, so do feel free to drop them in the chat during the session. And if either Pat or I are really good at multitasking, we'll try to answer them as we go along. Otherwise, we'll get to them during the q and a towards the end. If we're not able to cover the questions, please feel free to pass them along to your friendly Dataiku account teams, and they'll be happy to work with you on them. So without further ado, let's get into things and start talking about co build starting with why we built it. With tools like Cloud Code, codecs, anti gravity, what have you, access to frontier models, the promise of working with Gen AI, I think, is very attractive and very simple. Anyone can build whatever they want. So you have an idea, and it has never been easier to get to a working prototype, whether that prototype is an application, a workflow, or an AI powered solution. And AI coding tools have really changed not only the way that people who code for a living work, but also those who don't. I spend a lot of time using tools like the ones that I mentioned. Maybe you do as well, so I won't, I won't preach to the choir. But it's a very exciting and perhaps a daunting time, And it's fair to say that the way that you might be working today is very, very different from the way that you were working a year ago because of Gen AI and will be very different from the way that you'll probably be working in a year. So coming back to the promise of these tools that anyone can build whatever they want, this is also the big challenge of working with these tools. At their core, they generate code, and code doesn't scale across an organization unless everyone is good at it. So the people who understand the business best, your analysts, risk managers, operations leaders, domain experts, often can't inspect very easily or validate or approve what's being built, and much less also now at the volumes at which Gen AI lets people build things, solutions, and ideas. Second, what gets built is going to reflect the model's defaults. So not your organization standards. The model doesn't understand your business logic right off the bat, your governance requirements, your security policies, or even your operational best practices. Third, many of these solutions, because anyone can build now, are built outside of enterprise systems of record. So outside of approved data environments, governance frameworks, and IT oversights, we call this shadow gen AI. So while these tools really do help individuals move faster and faster in very interesting and exciting ways, they don't solve yet the broader challenge of collaboration that is essential to delivering trusted and governed AI at enterprise scale. And that's where co build comes in. So co build helps organizations turn intent and ideas into operational AI systems by combining speed and accessibility of conversational generative AI with the governance, visibility, and control that an enterprise needs. And our vision is quite simple. The idea is that really anybody should be able to build and deliver AI, not just a handful of specialists, and that the people who are closest to the business often have the clearest understanding of the problem that they're trying to solve and what good looks like. So they shouldn't have to wait in line for very limited technical resources every time that they wanna create an AI solution or create a solution. But at the same time, expanding participation can't come at the expense of governance. An organization still needs visibility into what's being built, confidence that standards are being followed, and the ability to maintain control as these types of activities scale. And that's the future that we see with Cobuild, more teams being able to contribute to AI delivery while AI and IT leaders can maintain full control, oversight, and enterprise standards. So Cobuild is an AI agent built into Dataiku, and it gives anyone, technical or not, the ability to describe an idea and generate an AI project from it. And what makes it different from most tools in this space is that the output isn't just code. It can be if you wanted to, but it's actually a full project win within Dataiku. So every step is visible. Every node is clickable. Your analyst, your operations manager, the risk officer can review what was built, review the logic, maybe catch what AI or the builder missed, iterate, etcetera, and do all of this in a shared environment. And meanwhile, co build is doing what it's best, understanding the problem, building the right plumbing for it, whether you're building pipelines that have a lot of ETL, whether you're building models, agents, web apps. And since you're doing all of this in Dataiku, you automatically benefit from everything that comes with using our market leading platform. So flexibility in choosing your infrastructure, leveraging push down compute, advanced analytics capabilities, and enterprise grade enterprise grade, excuse me, security and governance model, and the ability to build in a way that covers the entire life cycle of an AI project from design all the way to being ready to be pushed into production. And all this is happening in one place, so there's no porting things over from one system to another needing to rework. It's all happening in Dataiku already. And organizations that use Cobuilt will typically see value in three areas. The first is faster time to value. So rather than staring at a blank canvas and wondering, how do I begin? You can start really quickly with an idea and rapidly create a working solution. So Cobuild is going to take your description of what you wanna build. It can ask you clarifying questions if what you're saying is unclear, and it'll generate project components across data prep, machine learning, agents, and applications. So going from idea to working project is about to get much faster if you're not yet using Cobuild. Also, you may know exactly what you need to do on Dataiku and also know that it's gonna take you some time. And Cobuild can help you with that heavy lifting as well. So as an example, yesterday, I needed to change a data source in one of my projects, and I wasn't looking forward to updating a lot of the SQL queries because, of course, the column names weren't the same. So Cobuild helped me with the column mapping between the old and the new source datasets and then took care of updating my queries once I had validated that. The second big area of value is connecting AI to business users. Cobuild helps to lower the barrier to AI creation, allowing more business and domain experts to actively contribute to AI initiatives in a way that is aligned with their skill sets and the tools that they're used to using. So you might know what your business needs, what your team needs, but you might not know exactly how to build it. You'll just know if it's right or not. So you can build a flow with Cobuild and ask your teammates to review and challenge it. So Cobuild in this way expands who can create with data and with AI by making it easier to use the existing tooling on our platform. And third is governance by design. So Cobuild doesn't just generate code in an isolated system that is completely disconnected from everything else. It builds a Dataiku project with you. So every transformation, every action that it does is visible, auditable. Every step you can review and you can edit with governance, permissions, and security already built in. So that's an overview into the need that we saw within the market and why we started to build Cobuild, and now into what Cobuild specifically can do today. So as I mentioned, Cobuild is an AI agent. It has a set of skills and tools within a defined scope of capabilities that happen within a Dataiku project. What you see here covers the majority of activities that people normally perform on Dataiku. So Cobo can help you to find data, which is a very important starting point of a new project. Instead of having to reengineer all of your datasets, Cobo can help you find relative datasets on your instance. It can help you build your project. So whether you're building your pipeline visually, with SQL, with Python, building models or agents, Cobalt can also help you I'll say the word robustify for lack of a better term. So to make your project more robust, if it's on the path to being automated, So via scenarios, it can help you apply data quality rules, everything that you wanna put in place to make sure that your product is good and it stays good over time. And Cobo can also help you to document your project, either for future you, or for your colleagues to then review and then to iterate on. So what you see here isn't an exhaustive list of what Cobuild can do because that list would indeed be very long. We are shipping new additions to Cobuild once every three weeks. So this list is going to continue to evolve, and you can keep an eye on our reference documentation to see exactly what is coming out. But if you're using Cobuild, at a certain point, you might come across, something that it doesn't know how to do. And you'll end up with one of two scenarios. First is Cobuild is going to find a creative workaround, and you might learn something new about Dataiku that you didn't know before. My great example, I've been using Dataiku for almost seven years now. Thanks to Cobuild, I actually learned about the ordering capabilities in a concat aggregation of a group by a recipe, all because I just had the use case to do it. I described what I was looking for, and then Cobuild leveraged a group by a recipe. I thought that was odd. I leaned into it and, actually learned something new about Dataiku, which was great. The other thing that Cobalt might do is it might try and provide instructions. So it's trying to provide you with the answers to your questions. It's trying to help you. So it might provide DIY instructions for how to perform a particular action within the UI. This these are gonna be based on the model's knowledge of Dataiku. It might get a few details wrong, but it'll be directionally correct. And just know that if you do come across that scenario, we are working throughout the end of the year, maybe even beyond, to actually extend all the things that Cobuild can do. And Cobuild's current version is a small subset of all the things we would like Cobuild to be able to do, within your projects. Now the first thing you need to know about using Kobuild, now that we know what Kobuild can do for you, is that Kobuild is available on version fourteen seven, which was released almost two months ago now. And with version 14 dot seven, Cobuild and all of the other AI assistants are available to all of our customers. And all you need to do is to activate the AI services on your Dataiku instance once you've upgraded to this version or to a more recent version. I'll call out that we released version 15 yesterday. And doing this gives you access not only to Cobuild, but also to other AI assistance like the SQL notebook assistant, generate steps, generate metadata, AI search. These are some of our classic assistance that already exist on the platform. The second is once you've upgraded, is to choose how to operate our AI services. There are two modes. In the Dataiku operated route, Dataiku is going to operate Cobuild and the other AI assistants as a service for you. We'll take care of evaluating which models are the best fit for Cobuild, which ones perform the best, which ones are the fastest. And this is likely the easiest path to using Cobuild. Now, of course, this may not be a viable option for your company, and we understand that a number of our customers operate with very specific requirements. So we offer an alternative, which is to use your own LLM mesh connection. So with this option, you provide the LLM connection that's going to power Cobuild, and we've benchmarked how a variety of Frontier models perform on typical Cobuild tasks. We publish those results to our reference documentation to help orient you if you do choose this route to essentially bring your own LLM connection to Cobuild. Please do keep in mind that if your Dataiku instance is air gapped, we are working on extending support, for this type of environment by the end of the year. So to recap before I hand things over to Pat, we've talked about Cobill's value proposition, what it can do today, and a starter as to how you set it up. So, of course, if you have questions about what you've seen up until here, please put them in the chat or share them with your Dataiku account team. And with that, I will leave you in Pat's very capable hands. Okay. Hello, everyone. And, thank you very much, Ashley, for that great introduction to Co Build, what it is, what you can use it for, and how to set it up. I'm excited to talk to you guys a little bit about some of the things that you can do with Cobuild, and then I'm just gonna go ahead and and demo what it's like to to use Cobuild to build data science projects in Dataiku. So first off, I wanted to just, give you some inspiration of some things that you can do with Cobuild. So first off, Dataiku is a tool for, people to build data cleaning pipelines, train machine learning models, build custom AI agents, create data powered web apps, automate workflows, and a lot more. And so Cobuild is able to do all of these things too. You're just chatting with an agent instead of clicking around or coding these, artifacts yourself. So that's the number one use case. Number two is that if you're already a a Dataiku user, and you you have a portfolio of projects, I certainly have a number of projects that are sprawling to say the least. They have maybe hundreds of recipes. You know, it's been a while since I've touched them. So you can use KOBILL to refactor these existing workflows pretty easily, which is nice. And then the third use case, which is the the bane of many engineers' existence, is writing documentation. Kobuild is great at understanding a Dataiku data science project, and then writing a a really nice Wiki describing how it works. And so with all three of these use cases, what what's really nice with Cobuild is that you can collaborate with this agent. So you're not just letting this thing run wild and and build stuff and write documentation. You can edit all of these things yourself and go back and forth with Cobuild. Okay. So, Ashley, would you mind can I take over screen share? Great. Thank you. Okay. So let's walk through, Dataiku and using Co Build in Dataiku. And so I know some of you on the call have seen Dataiku, have worked with Dataiku before. That's great. A lot of these concepts will seem familiar to you, plus some of the new co build things. But I also wanna discuss a little bit of the basics of Dataiku for those of you on the on the call who haven't seen Dataiku at all, and you're like, what is this interface all about? And so I'm logged into my Dataiku instance right here, and this is collaborative. So, all of the Dataiku engineers, in my organization can can collaborate with me on these projects. And I've started out, I I wanna build a data science project to predict the, lifetime value of customers. And so imagine I sell I sell something through our ecommerce site. Right? And I have some data stored in Snowflake about, some customers. So I have an ID. I have their birth date. I have the price of the first item that they purchased, their gender. And then for many customers, I have a revenue total after one year. And I also have some web traffic information from, our our web team. So for each of our customers, I have their IP address, the number of pages they visited on our website, and whether they were a part of our our ad campaign, whether they visit our site through an ad. And then I also have a table of new customers. So these folks have have purchased one item from us recently, but, I don't know how much revenue we're gonna get from them within a year. So I wanna build a pipeline to clean up some of this data and join it together and train a machine learning model to predict revenue after one year, and then use this machine learning model to predict the future for these new customers. So I wanna understand who are my high value customers, who who is likely to spend a lot of money, at our company. So a pretty simple data engineering and data science use case, and we're gonna use Cobuild and some other Dataiku tools to build it. Okay. So, here's my project, and I'm gonna open up CoBuild. And I'm going to ask CoBuild to help me train a model to predict total revenue after one year. I'm saying use the customer revenue table as training data, the web traffic table for features, and then I wanted to make some predictions for the new customers table. And so CoBuild is going to first take a look at the data in this project. So you can you can see it previewed these datasets, and now it's starting to build some things. Now I've set the execution mode to allow automatic execution. So this means that I'm I'm letting CoBuild build things in Dataiku without me intervening, having to approve. We'll take a look at some of these other settings in a second, but you can see it's building some things pretty quickly. So first, it it decided to join these two tables together, and we can click in and see this was a left join. So we're taking our customer revenue table, and we're joining rows from the web traffic table where, this customer ID equals this other customer ID right here. And so this is a this is a nice benefit of using Cobyal Mid Dataiku is that, I can understand this without having to read mountains of code generated by an AI assistant. So it's a really nice experience to go in and click around and say, okay. I can I can understand what's what's going on here? Okay. So, it's doing those two things. And now it looks like it's creating a prediction ML task, trying to predict revenue total after one year. And so I know where these tasks live, and so let's take a peek at at that. Okay. So now CoBuild has kicked off a machine learning training competition. So it's decided to try out two different algorithms, a ridge regression and a random forest model. And we can see performance versus time it took to train. So higher r squared score is is better. So random forest performed the best. And if we click into this random forest model, we can understand visually how this model is is working. And so, for those of you who haven't done machine learning before or those of you who have, a random forest model consists of a number of decision trees, where we're taking an individual and we're, making a final prediction based on whether they were a part of the marketing campaign, the price of the first item they purchased, and some other features. Right? And so these trees are smartly built, through that training process. So this is how the model works. We can understand which features were most important. So, of course, the price of the first item that someone purchases is gonna be impactful in their one year end revenue. That that makes a lot of sense. It's like the one variable is relying on another. We can even look at, how the model performs across different subgroups. So when we're doing machine learning, especially as it relates to humans, it's really important that our model isn't biased toward one group versus another. So didactic use a lot of tools to understand how our model works across, men and women, which is, pretty interesting. And we actually have pretty similar performance across men and women, so that that looks pretty that looks pretty good to me. One thing looking at the scatterplot here, it looks like I had kind of have some outliers over here with high revenue customers. And so, you can feel free to ask CoBuild. So okay. So CoBuild came back and it says, okay. We've joined some data, trained some models, deployed the best model, scored. K. That's that's great. But you can go back and forth with CoBuild. So let's maybe say, like, how can we deal with high revenue outliers in the train set? So CoBuild is is not just good at doing Dataiku stuff. So building all of these things in Dataiku and understanding how to do joins and work with data in Snowflake and do machine learning, but it's also just a really good data science expert and then data engineering expert. And so you can chat back and forth with it to riff on ideas. Okay. And so we can say, okay. This is a nice statistics one. So you can log transform the target, rather than just predicting raw revenue. That's kind of a cool one. We can cap the target, segment the problem, so maybe have different models for high revenue customers and lower revenue customers. Yeah. I would start with a log transform chart. Okay. Cool. So, you know, Cobalt Cobalt is really good at just giving ideas about data science, whatever you're whatever you're trying to do. Let's try another thing. So, whenever you're building these pipelines, it's important to add in data quality checks. So these are live snowflake tables, coming in, and I wanna make sure that there's not, like, bad data going into my scoring or my training process here. And so in Dataiku, we can create data quality rules to, just make sure that certain columns look okay, certain values look okay. And Cobuild is great at doing that kind of stuff. Right? Some of these tedious data quality checks. You know? I, I I don't necessarily wanna spend all my time doing that. So let's take a look at what, what it did. So here I can look at this table in Snowflake, and you can see all these different tabs here. Right? And if I go to data quality, I can see that it created three rules, and let's compute them right now. And so it's checking making sure that all training fields are populated. So those are all the features required to, to train a model. Make sure that training revenue is nonnegative. Right? Like, negative revenue wouldn't make sense. We can also make sure that web pages visited is nonnegative. So these data quality checks make sense to me. If you ever wanted to edit these, you can just click on these, pencils anywhere. And maybe I wanna change pay pages visited to, like, 0.1 or something. I don't know. So something arbitrary like that. You could do that. So you can collaborate with CoBuild. Right? And then one other thing I wanna show you real quick is let's automate the retraining and inference steps. So let's say I wanna do model retraining monthly and, rescoring of of new customers weekly. Dataiku is really good at at automating these pipelines. You don't have to go in and manually click run buttons every week, every morning. Okay. And so here's a here's a nice cool little thing. Right? And so CoBuild needs some guidance for me as a as a human. And so sometimes it will ask you questions if it doesn't understand your, your request fully. So here it's asking, which schedule should I configure for retraining and scoring? Do I wanna use defaults monthly on the first at 2AM and weekly Monday at 3AM or use a different schedule? And a seems pretty good, so I can just click yep. Seems good. Yeah. And then it'll go and build some automations for me. Okay. Let's see. While this is happening, I wanna show you a bit, behind the scenes how Cobuild is working. Okay. It did create the monthly revenue model retraining scenario. So let's let's take a peek at that. And so here we can see that, monthly, starting on September 1 on day one at 2AM, we're going to build that training dataset. We're gonna run those data quality checks that we just created, and then retrain the the model. So that that looks really great. And the other scenario that we've created is weekly inference. And so weekly, Monday at 3AM, we're going to, again, build our, dataset that's ready for scoring. We're gonna run the data quality checks on that one and then score new customers, right here. Great. And we can even I, as a human, I wanna add to this process. Right? So maybe I wanna add, like, a Slack reporter. I wanna Slack our data science team saying this is done. I can add that right here. Great. And so, yeah, you know, what I mean, how long did that take us? Like, eight minutes maybe with me just kind of talking a lot over it, and we've created a an entire pipeline to clean and join this data, train a model, handle some outliers, create data quality checks, and automate retraining and scoring. It's pretty good in all visual. Okay. So let's let's go into the back end a little bit of Cobuild. And so for Dataiku administrators, you can set this up, across your entire user base, so all Dataiku users in your organization. And, Ashley mentioned earlier, you can choose to use the LLM operated by Dataiku, so Dataiku's OpenAI account, or you can bring your own LLM. So a lot of our customers are are enterprise large customers with large contracts with AWS and Snowflake and Azure and OpenAI and Anthropic directly. And so you can choose to bring your own LLM, and then just choose the model that you want as the main model behind CoBuild. So here we can choose Anthropic Cloud 4.8 Opus hosted on AWS Bedrock. We also have a fast light model option. So we do some model routing behind the scenes of Cobuild. So we can choose claud four six sonnet on Bedrock for that one. Yeah. So you can choose this. And as Ashley mentioned, we have a on our on our website for Cobuild, we have a list of the models that we recommend using as the base model. So we've seen good performance out of all of these models plus any any newer versions of these models. Great. And so a lot of companies also want to customize the way that these agents, work. So, if I'm an insurance company or an energy company or or a bank, I may want Kobuild to operate in some custom way that's that's unique to my organization. So, you can enter in custom global instructions in this field, which will apply to Cobuild everywhere used across an entire Dataiku instance. So I can say something like, I don't know, talk super concisely. We don't like too many words at x co. Whatever. Yeah. You can so you can customize how Co Build works for for your organization. You can also customize this at the project level. So if I go to my project settings and then co build, I can have project specific settings. So, in this project, we should only use visual recipes. No code ever. Right? So you can add that there, and then Co Build will respect that instruction. I can also set instructions for my profile. So for for me as a user. So I have a CoBuild tab, and I can say, I am a supercoder. Only use Python code for me ever visual recipes. Right? And so then Cobuild will only write code for me. It won't write any of the visual recipes. Some other cool things that you can customize for yourself are, Kobuild works in a number of different languages. So you can ask Kobuild to auto detect the language based on, just like what language I'm chatting in with it right here, or you can you can hard customize it to, to French, German, Gujarati, Thai, Japanese, whatever whatever you want right there. So it works in all these different languages, which is great, especially for our multinational, organizations and, yeah, others here. Cool. So I've shown you one of my three use cases of building cool things in Dataiku using CoBuild. Another cool use case is refactoring existing workflows. So this workflow is using a lot of visual recipes, and it's using Snowflake as the storage and compute underneath it. But what if I wanted to, let's say, can you change everything to Python code and store data in Databricks instead? And I'm gonna ask Cobuild to, ask me before doing any modifications this time. So I don't want it to just run wild and, create stuff without my permission. And so this here, I just wanna show you how in Dataiku, Cobuild will default to using visual recipes because no code is is really nice to to view, and collaborate on and not have to just read a mountain of code. But if you're a coder, you can absolutely code in Dataiku too using Python, R, SQL, Spark, your language of choice. And so Cobalt's, again, asking me a question. Should I create a parallel pipeline or replace the existing flow? I wanna create a parallel pipeline and not, I don't wanna remove my existing stuff quite yet. I wanna see them side by side. Right. And so I can use code in in Dataiku. I can also connect to okay. And so now Cobalt's asking me to, create stuff one at a time. So I'm gonna say yes. Don't ask again. Yes. You can create new datasets. That's that's good with me. And so in Dataiku, you know, we we can use code. We we can use visual. We can also connect to lots of different, data sources. So, like, we can connect to Databricks, Redshift, Oracle, s three, Trino, Azure Blob, Elasticsearch. Right? Any database that an enterprise needs to connect to. You can all have just in one single project, one single workflow. So you can see here we have, you know, we've got Databricks and Snowflake in one project, which is kinda crazy. Okay. Cool. And it looks like Co Build is doing some stuff. So it's, it's trying to copy our visual steps and okay. Something something failed. I wonder what happened. Let's take a look. We love these live demos. Okay. It looks like it's an it's a failed to authenticate to Databricks. Okay. And I think I know a good answer for this one. We love a live demo for these reasons. I wanted to we have a number of different Databricks connections on this, SE instance. So let's let's say, use the Let's use this one. I think it was using a different Databricks connection. But, yeah, this is a, like, a a great way to use Dataiku is for your non coders to use visual recipes alongside your coders using code recipes, especially when you have a use case that requires, multiple different connection types, like using Databricks and Snowflake and s three and and lots of other, options altogether. Cool. And let's take a peek at this recipe, this Python recipe that it's built. Okay. Great. And so if you remember, we we created a join from those initial two datasets. And so now we're just writing Python code to, to do a a merge. So, like, pretty simple pretty simple code here. And let's see if we can actually write to Databricks. Okay. Yay. We fixed it. So we're we're actually writing to Databricks now. So just as we can with Snowflake, I can take a look at this table in Databricks and okay. Yes. It does have both, columns from both tables in it. Yeah. So crazy. Taking data from Snowflake joining in Databricks. And so this will go ahead and try and recreate this all in Python. And so, yeah, Cobuild, really good for refactoring projects, also good at, coding. And then the last piece is documentation. And so in the same project, I can I can spin up another session of Cobuilt? And let's say, can you please write a Wiki documenting this project? Can you also add descriptions to all recipes and datasets briefly describing what they do slash r? Okay. Cool. And so this is the last use case that I mentioned. Right? So, documentation, always, miserable, and AI is is very good at at understanding what's going on in these projects and and documenting them. Okay. And while we, while I sort of pause and wait for co build to do its thing here, I wanna remind everyone, feel free to drop questions in the q and a, and we'll get to them in a in a second. I'm almost done with this with this demo, and then we'll we'll do, like, two more slides, and then we'll get to the questions. Okay. Great. So made a Wiki article. Really, really nice. So we have the purpose. We have the data sources of the projects. We have describing the main flow. We have our alternative Python, Dataprep branch. That's that's really nice. We have some data quality controls, and then we have the final output. So that's that's great. We have a Wiki. And then it's also added little descriptions to each node. So we can see right here, enriches training data with web traffic, trains first year revenue model, and then same thing with with some of the Python stuff. Great. So we have a a nice easy end to end data science project, code version and non code version. If we look at this one, this should be, this is funny where you it it decided to use NumPy's least squares, regression algorithm here, which is kinda funny. Okay. Great. Let's dip into some more some slides to close it out. Thanks, Scott. That was a great demo. Alright. So, we've been getting a couple questions in the chat. Thank you very much for asking those. I'll I'll review all of them also during the q and a and add some additional comments. But I figured we could also close out this session with a little bit of information about co build best practices. So we get a lot of questions around this. What is the right way for me to interact with co build? And, Pat actually showed a lot of these, best practices in action in his demo. So first is using project instructions or instance level instructions as steering. So just not steering in the codec sense of queuing messages, but just giving, Cobuild additional guidance as to what is the right way to do things, either for you as a user, for your team working on a particular project, or also perhaps if you're an instance administrator, how you would like Cobuild to interact with everyone who is using Cobuild on a particular instance. You can also definitely use separate chats for separate tasks. So Pat had multiple chats going at the same time, and all of those chats are visible in your own personal chat history. So you can maybe have one chat for all of the data prep stuff that you might be doing on one part of the project, maybe another chat for working with the scenario. And this helps CoBuild to work in slightly more incremental amounts, but it also makes it easier for you to review and to iterate on what CoBuild has done. Usually, if you write with any tool, a really large system prompt with a lot of complex instructions, one of the big challenges when your LLM or when your agent is done is then reviewing all of that output to make sure that it actually did what you had in mind, and then you'll have iterations. And it's easier to manage that work for yourself, actually, if those are spread across a couple different chats. So maybe that's not a co build best practice, but it's a person using LLMs best practice and just the limitations of human brains. Third is to review co build outputs, before you actually use them. So Pat took a look at the model report and looked at all the different capabilities that we have there to help you understand what the model does and how it calculates certain things. That's certainly a very good best practice practice to apply normally and, certainly, anything that is Gen AI generated, to make sure that the models and your agents are doing exactly what you wanted. Because sometimes, they might do something, and you'll realize that, actually, that's not what you wanted, that you didn't express it correctly, because you probably didn't completely understand what you want. And so sometimes having an agent build a prototype for you and then actually in reviewing the prototype, you better start to understand what you wanted in the first place. Number four is to use code selectively for simple flow tasks. So there might be some things that are simple enough that actually just clicking in the UI and doing it yourself might be a little bit faster. So things like deleting a dataset or renaming a flow zone. Those types of things, you could probably do them directly, and my advice would be to really use Cobuild for, a lot of the heavy lifting tasks. So as an example, if you wanted to create 20 data quality rules on your dataset, Cobuild would be a great a great tool to use. Or if you realize that you have to make a change somewhere, but you have to make the change in 20 different places, Cobuild is also great for that. My personal favorite is, using Cobuild to rename datasets in old projects based off of what the datasets kind of mean within their spot in the flow. It makes a lot easier to read a flow and understand what it's doing. Fifth, I'm having trouble counting, is to plan before you have Cobuild execute a very large set of tasks. So you saw Cobuild ask Pat a clarifying question. You can also ask Cobuild to kind of go through an iterative loop with you, so to discuss how you might approach a problem, to choose a particular way to do it, and then to have Cobuild explain how it would build it for you. And that is an option that you currently have. You can then, if you want, save that plan to a Wiki article or save that plan to a flow zone if you're working specifically within a flow zone. And that'll also then help you after to check the work that Cobuild did. And I'll also note that we're working on adding a plan mode to Cobuild so that it's an explicit planning mode. Cobuild then has that plan, and then can also work with that. And then fourth is to work in reviewable blocks, just so to make it easier based off of all the different things that Kobold can do in a project. So it's easy for you to check the work that Kobold is doing and to make sure that what it did is aligned with, what you had in mind for what you wanted CoBuild to do on your behalf. So with that, these are just a couple best practices that we're seeing are making a big difference in how people are using CoBuild. I'll turn now to questions, and I'll start with some of the questions in the chat, just, reviewing some of the, questions that have been asked. Thomas asked a question about what is the difference between, since I did mention that 15 has just dropped, what we see in 14 dot seven three, which is the most recent version of fourteen seven, and what's in 15. Pat actually showed you what that major difference is, and that is the ability to provide a user level prompt for Cobuild so that you can go modify in your user profile. And that gives Cobuild a specific set of instructions just for you that Cobuild is going to bring with it across all of the different projects in which they you might be working, with Cobuild. Perhaps a couple other notable updates between 14 dot seven three and fifteen is the fact that, when you view your history, these are small usability updates. When you look at your conversation history, Cobalt is going to auto name those conversations to make them a little bit more descriptive and then easier to pick out the one from a list of conversations. And you can also pin conversations. Now we know that 15 is a major version, and that a lot of our customers take time to upgrade to 15 because there might be some technical blockers for making that upgrade. So we'll actually continue to release co build updates through the end of the year on a 14 dot seven dot patch release. So that if, for whatever reason, your company isn't ready to upgrade to version 15 yet, you'll still be able to benefit from the latest and greatest cobuild improvements, on a version fourteen seven. So as an example, in a few weeks, we'll be releasing version 14 dot seven dot four. So even if you are on version 14 dot seven and that's where your company is gonna stay for a while, you'll still be able to benefit from a lot of really interesting things that Cobalt is going to do, notably things like plan mode or Cobuild being able to interact with project standards in a project. So little answer for Thomas. Thank you for the question. Let's see. Christine's question. Without Cobuild, how do we do documentation? So, when Pat asked Cobuild to document the project, Cobuild built a Wiki. Cobuild also added short descriptions to the elements in the flow. That would be how you would one way that you could traditionally document your project. You could write that Wiki article. You could add comments to objects in the flow. If that flow were organized into flow zones, you could then also add descriptions to the flow zones to make them easy to read. Flow zone descriptions, you can now also display on the flow. So whatever description of a flow zone you want is actually just gonna directly display on the canvas of the flow. So that would be one way that you could potentially document your project within Dataiku. You could do that yourself, or you could also ask Cobill to do it. What we also see some customers doing with Cobill now is, typically, documentation has some format of a template, would actually be to have a template in a Wiki article and then ask CoBuild to write documentation as a Wiki article following a given template. So CoBuild can then also read the Wiki articles in a project, use a template, and then generate a template compliant, Wiki article for you. Let's see. Is there Pat, is there any question you you'd like to answer? Sorry. I'm typing out answers to some of the questions in the chat. Maybe we should cover the I in the q and a, Yes. I think, It's, on the top. we, like, we can't, like, type out answers in the q and a. So maybe let's cover the q and a questions. Sure. What do you think? Sure. Can I okay? So one is, can I use Cobuild to generate API endpoints and queries? Do you know the answer to that, Ashley? I actually don't know off the end. I don't believe currently Cobuild can publish an API endpoint. So if we're talking about anything that is built within a project, our ambition is that Cobalt will be able to do that. But Cobalt doesn't know yet about how to deploy things. So if you're deploying a project to a production node, deploying something as an API endpoint, Cobalt can't yet do that. But if that is a use case that you feel is really relevant to the daily work that you're doing on Dataiku, feel free to reach out to your account team, to talk about what you have in mind, and then they'll relay that ask, to us on product. Dataiku, thankfully, makes that pretty easy. Do you want me to just show on the screen real quick? Why not? Sure. Okay. Can folks see this again? Okay. Yeah. Great. Yes. Okay. So I asked it. So you can you can always ask Cobuilt what what it can and can't do, and we'll definitely be adding this. But if you have your trained model right here, you can click create API, Call it, like, revenue API service. Call it predict revenue. And then you can customize this a little bit. So you can add, like, security layers to it. You can add, like, OAuth authentication or whatever, but it's super simp I just click publish to Deployer, publish, and then here I have a okay. Signing in. Classic. Then I just hit deploy to, and then I can deploy it to, like, Kubernetes or SageMaker or, wherever. So even though it Cobuild should be able to do this soon, but it is just a few clicks right now. Great. Thanks for showing that. Yeah. I'll queue it up with, Vikrant's question. Any plans to add an option to configure the LLM used in co build at the project level in addition to the current instance level setting? And, yes, I'll echo, Pat's response that we don't yet have plans to do that. But if that's a use case that you have, please don't hesitate to reach out to your account team with that feature request. A question from Casper. What language is best to use? LLMs are best trained in English. Is this the case here as well? That's a great question, and the answer is going to depend on the Frontier model that you're using to power Cobuild. Again, all of those capabilities are really limited by what the model underneath can and cannot do. With AI services operated by Dataiku, we tend to pick the latest and greatest Frontier models to to run CoBuild, so it works reasonably well. And the languages that you see or that you saw on the drop down correspond with, what the current set of models we're using OpenAI. We recently upgraded to five six, correspond to what those models do do pretty well. But, again, that capability is gonna be based off of what the LLM that is powering Kobuild is good and knock it out. And most of these LLMs tend to be much better in English, than other languages. But we did want to provide the ability to customize Cobuild's language because Dataiku, you know, even if we look at where folks are coming from today, coming from all over the world, and if there's an opportunity for us to make, Dataiku easy to use, often that is actually in the language that you see on the interface. And LLMs are a great way, to essentially get over some of the hurdles of, Dataiku being in English, in French, and in Japanese. It's great if you speak those languages, but a little bit more challenging if your native language is something else. Should we tackle the admins disabled co build question next? Yeah. You can take a crack take that? one. Yeah. Casper says, in my firm, the administrators disabled co build, probably for a reason. What can go wrong? Is it in the design or in the running phase? Can it use data from other developers projects without permission? What can you do about it? I think it's a great question. Dataiku is used at many highly regulated IT security conscious, large enterprises, and we found that a lot of our customers just take a little while to enable and adopt new AI features, whether that's Dataiku co build or anything else, whether it's just using an LLM period or using another tool like Cloud Code or something. And so I wouldn't read in too much into your administrators maybe disabling it. I think it's worth asking them if they're currently evaluating Cobuild. We're definitely working with a lot of customers where they have an an IT sandbox where they just like to kick the tires with especially new features, the new AI features before then rolling it out to the the larger Dataiku community. And so it's possible that they're doing something like that. Our customers are security conscious. They wanna test things out before rolling them out to all users. But that being said, you shouldn't have any security concerns with Cobuild. Cobuild operates with your individual user permissions, So it's not going to be able to read any data that you, as a user, cannot read already. And the same thing with projects and models and things in Dataiku that it'll only be able to access what you have access to. Alright. Well, it looks like we're almost at time. So I'll very quickly answer Cara's question. Will the recording of the webinar be emailed to registered attendees? Yes. You'll get both the recording and also a copy of the slides. And since we're almost at time, we won't have time to get to everyone's questions. I'm sorry. We've tried to answer them in the chat as best as we could. But do feel free then to contact the folks on your Dataiku account account teams. They'll be more than happy to answer your questions with respect to Cobalt and to help you get started using it. My last sign off, plug is to mention our Dataiku succeed conference. So if you're interested in learning more about Cobuild but also about all the different new capabilities that will be in Dataiku come September. Please join us at Dataiku succeed that is gonna be in New York on September 24. All you need to do is scan the queue. Oh, I need to be sharing my slides. Whoopsies. Scan my verbal QR code. Yeah. Yes. Scan scan this QR code to sign up. We're bringing together leaders, who are doing a lot of the hard work of translating these ambitions that, that we have with AI into measurable business impact and outcome. So you'll hear directly from leading companies about their experiences, learn about what best practices are and the choices that they're making about, you know, people, governance, orchestration, and technology, and how all of those things come together in Dataiku. So with that, I will say thank you very much, for joining us for this hour today. I hope that, we were able to give you a very thorough introduction to Cobuild, so you understand, what it is, what it can do for you, and have some ideas about how you might use it in the work that you're doing with Dataiku. With that, I'll sign off, and wish everybody a very happy rest of your Thursdays. Thank you, everyone. Thank you so much. Bye. Take care.