Building Enterprise AI Agents: Repsol's Genie Code for Predictive Analytics

Summary

Repsol built a multi-agent platform on Databricks that orchestrates the full ML lifecycle, combining a Knowledge Assistant for domain expertise retrieval, Genie SQL for business-language data access, and Genie Code for generating production-ready predictive model code using Repsol's internal libraries.
The platform addresses the core bottleneck that most practitioners master only one or two of the four required dimensions — business knowledge, data access, modeling skills, and software best practices — by assigning a specialized agent to each role.
Repsol's first digital wave (2018–2022) delivered 500+ AI use cases with over 1 billion euros of economic impact; the second wave from 2023 adds generative AI as a core strategic element, with the multi-agent Genie Code platform as a key deliverable.

Building Enterprise AI Agents: Repsol's Genie Code for Predictive Analytics

Building predictive models requires mastery of four key dimensions: business knowledge, data access, modeling skills, and software best practices. Most practitioners excel at one or two but lack depth in others. This creates bottlenecks in industrializing machine learning across business units.
In this talk, Victor Vaquero Soto from Repsol demonstrates how a multi-agent platform built on Databricks orchestrates the full ML lifecycle. Learn how the Knowledge Assistant retrieves domain expertise, Genie SQL bridges business language to data, Genie Code generates production-ready code using internal libraries, and an MCP server ensures quality at scale. See how this architecture powers Repsol's soft sensor use case, delivering predictive models for refinery operations.
🤝

Chapters

FAQs

What is Genie Code and how does Repsol use it for predictive modeling?

Genie Code is a Repsol-built AI agent on the Databricks Data and AI platform that generates production-ready machine learning code using Repsol's internal libraries and software best practices. It is one component of a multi-agent system designed to automate the full lifecycle of building predictive models, particularly time series forecasting for refinery soft sensors.

What is a soft sensor and why does Repsol use predictive models for refineries?

A soft sensor is a predictive model that estimates a physical measurement — such as temperature, pressure, or composition — from other available sensor readings, reducing the need for costly physical instruments. Repsol uses time series forecasting models as soft sensors in refinery operations to improve monitoring accuracy and support real-time operational decision-making.

How does Repsol's multi-agent architecture address knowledge gaps in ML development?

Repsol's multi-agent platform addresses the challenge that most practitioners lack expertise in all four dimensions needed for production ML — business domain knowledge, data access, modeling skills, and software best practices. The Knowledge Assistant retrieves domain expertise, Genie SQL bridges business language to data, Genie Code generates model code, and an MCP server ensures code quality and adherence to internal standards.

What has been the business impact of Repsol's AI digital transformation?

Repsol's first digital wave from 2018 to 2022 launched more than 500 AI use cases involving over 1,200 people and delivered an economic impact exceeding 1 billion euros, along with significant CO2 emission reductions. The second digital wave, running from 2023 to 2027, continues this focus while adding generative AI as a core strategic element alongside the multi-agent Genie Code platform.

Full transcript

[00:09] Okay, so let's start. So, first of all, let me thanks the data bricks for organizing this really amazing event and giving me and Repsol the the opportunity to be here today. And I really hope you enjoy the the presentation. So, this is the title of the of the presentation,
[00:24] scaling at a scale, the journey of Repsol artificial intelligent product, what we call right, and Genie code at Repsol. And my name is Victor Vaquero Soto. Okay, so first, let me quickly walk you through the agenda. So, first, we will
[00:41] start with a short introduction about me and Repsol. After that, we will explain how Genie I and foundation models are changing the way we will predictive models. After that, we will explain our approach and architecture behind it.
[00:57] Then, we will see a demo about the product. And finally, we will finish with the next steps of the product that we have been working on. Okay, so let's start. Let me introduce introduce myself in a
[01:12] bit more detail. As I said, my name is Victor Vaquero Soto and I work as a senior data scientist at Repsol, where I lead the predictive analytics team, mainly focused on time series forecasting. So now, let's talk about Repsol, a company that many of you probably
[01:28] already know. Repsol is a Spanish multi-energy company based in Madrid with people and business all around the world. We have more than 24,000 employees and we are divided in five main divisions: E&P, industrial
[01:43] transformation, client, low carbon generation, and corporation. And our objective, our goal, is very clear, to put the customer at the heart of everything we do, while becoming a S-Zero company by 2050.
[02:03] So, regarding our digital journey, we have been working to become a truly data-driven company, and this is something that extends to the highest levels of our leadership. So, here you can find some words from our CEO about this process. And to make this happen, we have
[02:19] launched two different digital waves. The first one, between 2018 2022, where we launched more than 500 digital cases using artificial intelligence with an extreme economic impact of more than 1 billion of euros, with more than 1,200
[02:37] people involved, and of course with an extreme impact in CO2 reduction emissions. And then, the second digital wave, which started in 2023 and runs until 2027, where we continue focus on artificial intelligence and data-driven culture,
[02:54] but we have also added generative AI as a key component, as a key part of our strategy. So, uh over last year, we have been working and developing the components that right now allow us to scale and accelerate.
[03:11] So, on the data and compute side, we have our platform, platform, sorry, called Aria, where all of our data is stored and governed, and where we run all of our artificial intelligence processes. And on top of that, we have a set of
[03:26] tools, a set of libraries, called RIPE, Repsol Artificial Intelligence Products, products. They are fully integrated with our platform, with Aria, and we can use these libraries to work with different type of problems, like anomaly detection,
[03:42] mathematical optimization, computer vision. And the last one, Repsense, this is the library that I lead. It's the library that we use for time series forecasting and is the library we are going to discuss in a bit more detail in the next minutes.
[03:59] Okay, so let me introduce you Rep Sense. So, Rep Sense is a Python library specialized for time series forecasting. So, this library is fully documented and deployed with continuous integration continuous deployment. So, that means that it can be used across the whole
[04:14] organization, the whole company. And this library has been designed to be used in two different levels. The first one by our data scientist team and our providers that can use the full capabilities of the library to build and deploy models.
[04:30] And also from the business user side, business user with an strong knowledge about the data with we call subject matter experts which can use the the library uh in a low code way with low code interfaces. So, that means that they
[04:46] don't need to programming. Okay, so now let's first let's focus on the data science side. So, we have applied the library across different digital cases of our five main divisions. So, for instance, we have used it for fuel demand forecasting in
[05:03] our refineries and in our distribution network. We have also used it in supply chain demand planning and for instance, the last one in demand forecasting for pricing strategies like in our client and 3D visions. And we have working
[05:18] using this library on but we had managed to deliver value of our to our business. So, the ideas with this kind of libraries, we have managed to a standardize, accelerate and reduce the cost of model development. Okay, now let's talk
[05:34] about business. So, we have a really nice real business case in our refineries. So, this digital case is called soft sensor and our business users from our refinery use this library, use RepSenz, to estimate properties
[05:51] about our refineries like composition, product quality, those kind of things, in order to keep specifications under under control. So, we have more than 50 users from all of our seven refineries, and
[06:07] more or less we develop between 10 20 models per per year. And in this digital case, we follow these steps. So, the first one, the expert identifies the need. The second one, data capture. The third one, modeling and
[06:23] configuration of the model. The fourth one, validating the results. And finally, deploying into production the the model. And it is worth mentioning that steps number two, three, and four can be done either by data scientist or
[06:40] by business users. So, at the end, I think this approach is is really nice. We have deployed and built a lot of models, but we have also detected some pain points that I would like to discuss with you. So, the first one is
[06:56] that when business build the model, they depends on data scientist. At the end, they are not experts regarding modeling, so that means that data scientist teams needs to validate the metrics, the models. Also,
[07:13] the code that the business users develop needs to be in a lot of cases refactor because they are not expert programming. And we have the same problem in the other direction. So, when data scientist build the model, at the end, they are not refining and chemistry experts. So,
[07:30] that means that we have a huge learning curve, and of course, we have a strong dependency with the business because we need to validate which variables we use, and the metrics of the models.
[07:47] Okay, so now let's discuss how generative AI and foundations models are changing the way we build this predictive models with the idea to build better and easier predictive models like in this digital case. Okay, so I think a lot of things has
[08:05] happened over last years. So, first we have the large language model boom with the launch of ChatGPT in 2022. Then we have the arrival of 10 series foundation models, being 10 GPT the first 10 series foundation model in
[08:20] 2023, and more recently we have the rise of agentic AI and the concept of by coding. And our idea is our discussion is how can we combine all of these recent advances with all of the knowledge that we
[08:36] already have in our company with the idea of, of course, build better predictive models. Okay, let's start discussing generative AI. I think here the idea is really simple. We have a user, the user prompts or ask
[08:51] a question related to the model to a large language model and after a few seconds the user gets back the code for the modeling without any coding needed. So, here we can imagine our subject matter expert from our refineries asking something like, "Okay,
[09:07] build a forecasting model for sulfur content sorry from from whatever." And after a few seconds they get the code the code sorry. So, what does the what is the result of this approach? I think a really really really really happy user because every
[09:24] time they press the button, they get the code, they get the model. So, I think it's really easy and for this reason the user I think is always happy always happy sorry. But, I think I would like to discuss this approach with you a in a
[09:40] bit more detail. Of course, this approach has a lot of advantages like it allows a radical democratization. So, that means that the domain expert can build models without coding skills.
[09:55] So, the second one is allows us a faster prototyping. And the last the last one, sorry, it lowers the barrier to iterate, experiment with different use cases. But, this approach is also has some hidden risk that I would like to discuss with
[10:12] you. So, the first one is the code we get does not follow best practices. So, we get code that is untested, undocumented, and unversioned. There are poor analytical practices like data leakage, wrong validation, and
[10:27] overfitting. And the last one is a code that is hard to industrialize. That means that works locally, but typically breaks into production. So, I think 10 years ago everyone agreed agrees that with bad data we get bad models. So, now
[10:43] the question is can we really expect good models just from prompting? Because getting something working is really easy, but getting in good with good conditions, good code quality is is still really really really hard. So, the question we
[11:00] ask ourselves is, okay, can we use this by coding this generative AI approach truly useful in our industrial environments with the idea of building good models? Okay, now let's move to the second to
[11:17] the second point. The arrival of foundation models for time series. So, these models are pre-trained on billions of data data points from diverse time series like finance, energy, weather, those kind of things. And they are models that are
[11:32] capable of zero-shot forecast forecasting of unseen data without any retraining. So, we already have a foundations models from big companies like Chronos from Amazon, Moirai from Salesforce, or TimeGPT from Nixtla among others.
[11:49] And here the the idea again is really really simple really simple. So, we can imagine our user user from our refineries asking something like, "Okay, so forecast the next 30 days of sensor whatever you you can imagine." They called the foundation models and
[12:05] after a few seconds, we get the forecast. And again, without any training and any coding skills needed. And the result of this approach, so you can imagine again again our user that we discussed a couple of minutes
[12:21] ago. And again, we have a really really really happy user because every time they called the foundation models, they get a forecast. And again, we don't discuss that this approach has a lot of advantages like rapid prototyping, zero-shot forecast in
[12:38] minutes without any training needed, no time series expertise required for the subject matter expert. And the last one with baseline, so we can get really nice results just in a couple of minutes. But again, this approach
[12:55] has some risks and limitations that we cannot ignore. So, the first one is data leakage. So, maybe the training data does that these huge companies use for training may overlap with our evaluation data. The second one, the domain gap. So, in
[13:11] our industrial processes, we have very unique and very rarely dynamics that is rarely seen in public training data sets. And the last one, these models are black box models. So, that means that we don't have control over feature engineering,
[13:28] validation strategies, or model interpretability. And here again, we the question we ask ourselves is, okay, can we exploit this technology, these models in our industrial time series, still following our best practices?
[13:45] Okay, so let's discuss our our approach. So, when someone builds a predictive model, whether they are data scientists or business user, we think that we need to master four key dimensions at
[14:00] the first time. So, the first one, business and process knowledge. The second one, data access and knowledge of the data. The third one, model development skills. And the last one, the fourth one, software best practices.
[14:17] And the reality is, most of the cases, whether they are data scientists or business user, they don't master the four key the four dimensions at the same time. So, data scientists, they know very well how to build models, and of course, they follow
[14:35] good software software practices, but they don't have knowledge about the business. And from the business user side, they know a lot about the business and the process, but they don't have programming skills.
[14:50] So, our idea is to build and deploy an agent that master the four key dimensions at the same time, with the idea of orchestrate the full machine learning life cycle. So, the idea is that this agent understand the process contest, connects
[15:08] the different data sources that we have, use our internal libraries, like Repsych and finally ensuring good analytical practices, like avoiding data leakage and running validation checks. So, as a summary, the agent
[15:23] orchestrates the full machine learning life life cycle while filling the knowledge gaps. Okay, so let's explain the architecture we have uses to deploy this agent and the reason because we we choose
[15:40] Databricks for that. I think there are three main reasons. So, the first one is all of our data is a store and governed by Unity Catalog. So, it was the natural choice. The second one, Databricks is our main platform where all of our teams works
[15:57] daily. So, our data scientists uh typically deploy in Databricks. And the last one, Databricks offers a really really powerful agent platform. So, where we can build different types of
[16:12] agents. So, here on the bottom you have the four agents that we have used for this machine learning orchestration. So, the first one is the knowledge assistant. We use rack technology. The second one is Genie, the SQL agent
[16:28] for data access. The third one, Genie code. The agent responsible for generating code. And the last one, the CP service that allow us to connect to external external tools. And orchestrating the
[16:43] communication between all of these agents, we have the supervisor the supervisor agent. So, now we are going to discuss all of one of all of these agents one by one. Okay, let's start with the with the first one.
[16:59] The knowledge assistant. So, this is the agent that we use to have access to the process and business information. So, in our case, we have loaded operation manuals from our refinery, process descriptions, and technical documentation to the system. All of this
[17:16] information go to a semantic layer. So, here we use semantic chunking with overlap. Then, these chunking are converted to to embeddings. In this case, we use the Databricks GT large English model. And finally,
[17:34] these embeddings are stored in a Delta vector search. So, every time a business user or developer ask something related to the business, the system retrieves the more relevant contest and passes the question to a
[17:51] large language model. So, this is the first agent that we use in this Subsensor Digital case. And of course, the evaluation of this agent is fully integrated with MLflow, the tools that Databricks offers for for those kind of things.
[18:07] Okay, now, let's talk about let's talk about the second component, Genie, the SQL agent. So, here the idea is again really, really simple. The user ask for data in natural language and Genie handles the rest. So, in our case, we have a lot of
[18:25] information from PI systems. So, here we mean sensor data from our refinery units, variables like temperatures, pressures, flow rates, control variables, those kind of things. And all of this information is a store and
[18:40] govern in Unity Catalog by more than 100 tables and thousands of sensors. So, here the idea is that Genie use all of this information like data and metadata
[18:55] to generate automatically the query that allow us to access to the to the data. So, let's again imagine our user from our refinery asking something like, "Okay, show me the average temperature of unit whatever during last week." So,
[19:12] the user ask this question to Genie, and after a few seconds, we get back the query. Okay? And what makes this really, really nice is the combination of Genie SQL with the rack. Because with this combination, we generate a semantic
[19:28] layer. So, that means that a variable a Russian sensor like 612TE0813 becomes something like the temperature of the at the outlet of the unit 612. So, for this reason, this combination is
[19:43] really, really nice because you can establish relationships between raw variables and process knowledge. So, as a summary, the the user focuses on the question, not the query. And with the combination of Genie and rack, we
[19:59] bridge the gap between domain language and the raw data. Okay. So, now now let's talk about Genie code, our developer agent, the one responsible for for generating code. So, here the idea is that we teach Genie
[20:16] code to use our internal libraries, and Genie code acts as an AI guardrail, so ensuring best practices in model development and a smooth path to production. So, how did we teach this to the agent? So, we use the skill concept.
[20:33] So, that means that each functionality of the library becomes a skill. So, in the skills, we have three different files. The first one, the skill markdown. So, we use the this file to teach the agent what the skill does and when to use it.
[20:49] The second one, the reference markdown. So, here we have library docstring and API details, API documentation. And the last one, the example markdown which uses examples with real code patterns. Okay. So,
[21:04] every time a business user or data scientist developers ask something like build a soft sensor model for furnace outlet temperature using unit blah blah blah Genie code follows four different steps.
[21:20] So the first one reads and loads the skill. The second one plans the pipeline. The third one generates code and finally executes and validates the code. So our results of this framework we have an output. In
[21:38] this case it is code that is using our internal libraries. So the idea is this skills acts on AI guardrails and the agent generates and executes code using our entire internal libraries. So that means that that means sorry that we are
[21:53] ensuring best practices. And last but not least we have the MCP. So in this is in this case the for the MCP deploy we have used data bricks
[22:09] apps. And here this MCP we have built a peer review of code. So that means that we have a set of tools that analyze the code. So every single piece of code generated by Genie code
[22:25] goes to this MCP server. We run different tools. For instance here you can see flake eight which analyze Python style code. We also have vulture for this code detection or the last one we also measure
[22:41] the internal libraries usage. So that means that we can analyze and track the quality of the code that Genie code is generating for for us. And finally we have built a dashboard using again
[22:57] data bricks apps and streamlet. And over this dashboard we can compare, analyze, and monitoring the different KPIs. So, as a summary, every line of agent-generated code is audited, scored, and tracked.
[23:17] Okay, now let's turn Let's move to the to the demo. So, let's see this product in in action. Okay, so here we are in our Databricks workspace. So, as you can see, we we are inside the supervisor agent section. So, here you can find our supervisor agent. In this
[23:33] case, we call it SoftSensor Advisor. And if we go inside, just wait a couple of seconds. Yeah, here you can find the different sub-agents. So, the first one is GeniePI. This is the Genie SQL agent.
[23:49] Remember, this is the one responsible for the taxes. The second one, the Process Knowledge Advisor, we use RACK technology. And finally, the the MCP Review, connecting with external tools. And the supervisor agent has a has a set of instructions that tell it
[24:06] how to work. Okay, so now let's test it. So, imagine that we are a business user from our refinery, and we want to build a predictive model for a property, in this case, a penetration property. So, the first question the user ask, "Okay, tell me the most important variables." You can
[24:22] see that this question goes directly to the supervisor. The supervisor detects that it's a question related to the business, and passes the question to the knowledge assistant, to the RACK system. So, as you can see, after a few seconds, we have the
[24:40] the answer of from the from the agent. In this case, we have the most important variables for this penetration property. We can also validate from the knowledge assistant the original information, so to ensure that the answer of the agent is correct.
[24:56] And as we said, from this rack, from this knowledge assistant, we can also get a semantic layer. So, that means that we can establish relationship between raw sensor and processing information. So, here you can see that
[25:11] raw sensor like 612 temperature becomes something that it has a physical meaning. And after that, you're going to see a summary from the agent from the agent, sorry, with the most important variables that we should use in our modeling
[25:26] strategy. And of course, some modeling inputs, some modeling tips that are really nice when we move to the to the modeling. So, in this case, you can observe some important modeling consideration and next steps for the for model development that are really,
[25:42] really, really nice. Okay, so once we know the most important variables, the next thing is, okay, give me the data. So, as you can see, we ask the to the supervisory agent the query to to get the data. So, again, we send this
[25:58] question to the supervisory agent. In this case, the text that is a question related to the data and passes the question to the Genie SQL. So, after a few seconds, you're going to see that the first thing that Genie SQL is is performing is validating from the
[26:14] original information which variables we have in our systems and which one we don't have. So, it's really nice in order to validate information. It's also validating the target name. So, in this case, remember that we want to model a property called penetration.
[26:29] And finally, I after a few seconds, you're going to see the SQL query that we can copy and run in a notebook with idea of getting the the data. So, here you have the the query. And what is really, really nice, also Genie give us
[26:46] some data cleaning tips from the documentation like natural or normal values our uh properties. Okay, so really, really nice. Now, we move to a notebook, so where we have already loaded the data with the previous query that we get a couple of
[27:02] seconds ago. Here you have GenieCode. Remember, this is the one responsible for generating code. So, inside the settings and the configuration of GenieCode, first we have the NCP servers. Remember, here we have some tools
[27:17] for analyzing code quality. We also have instructions at the workspace levels and and the user level. So, it's really nice because we can combine different configurations. And finally, we have the skills. So, remember, the skills is the
[27:33] concept that we use to teach GenieCode to use our internal libraries. So, here you can find different modules from our libraries. We have the three files we commented a few minutes ago. And here we are we are inside the modeling skills and you can
[27:49] uh observe some tips that we we give to the agent like feature selection strategies, good modeling practices, and those kind of things. Okay, so let's test GenieCode. So, first thing we ask to the agent, "Okay, I want to build a model. Please tell me what
[28:06] are the steps to follow?" So, you can see how Genie is reasoning, how it's loading the skills. First things that you can see that it's making use of our internal library. In this case, uh repsense. So, here we have the
[28:22] typical workflow that we we follow when we build a predictive model. Okay? So, now once we know the workflow, first thing, "Okay, please run an analysis of the target distribution." I think this is the first thing that you used to perform when you start uh an
[28:38] analysis. So, here again, you can see how GenieCode first is analyzing the code that we already have on the notebook. It's loading the relevant skills and after a few seconds, please on the left observe that Genie is going to start to
[28:55] inserting code. So, here you can observe this code. It's making use of our internal library Rep Sense and also runs the code. So, here you can see the results of this target analysis. Here we have some typical plots like
[29:12] distribution of the target, the time series plot and what is really really nice, Genie code takes these results and analyze it. So, here you can observe some conclusion that the target follows an uniform distribution and some
[29:28] best practices that we should follow in the modeling strategy. So, here you can observe some recommended model strategies. Okay. So, we follow the recommendation of Genie code. So, we want to model the full range uniformly and
[29:45] we move to the next step. So, imagine that now we want to select the most important variables. So, we ask Genie for that. So, here you can observe the the question. I want to select the most important variables. Again, Genie code first load the skills,
[30:01] analyze it. So, in this case we are in the feature selection step and after a few seconds, again on the left, the code is going to appear. So, here at the beginning you can observe that it comes from our library, the module. Genie code runs
[30:18] the code and here you have the results of this feature selection approach. After we have these results, again we move to the right. Genie code is analyzing these results and you can see how the raw sensor names that we have on the
[30:34] left has a physical meaning of the on the right. So, Genie code is making use of our internal documentation and it gives us an physical interpretation which is really really nice. And finally the last step we ask Genie
[30:50] to train a a predictive model. So again Genie code in this case is going to load the modeling skills, is going to analyze it and finally again on the left the code in this case for modeling is going to is going to appear.
[31:06] So let's wait a couple of seconds, yeah. So again it's going to edit in the the notebook and inserting the the code. Again at the beginning you can observe that this regression component comes from our
[31:21] library. Here we have in this case Genie code decided that the best model is XGBoost that I think is really nice because in this case is a non-linear problem. Here we have the results from the modeling and of course we have some metrics. So here you can observe the median absolute error for instance and
[31:38] what is really really nice Genie code is analyzing the metrics. So in this case we could think that the model is okay but in this case you can observe that Genie is interpreting that the model has an over-fitting which is true. It also analyze root causes and what is
[31:56] really really nice you can observe that recommended next steps in order to avoid this over-fitting. And finally we close the workflow in this case analyzing the quality of the code the code sorry. Remember we have MCP
[32:11] peer review. In this case we run it auto manually but we can also run it automatically. Here the Genie code is calling the different tools from this MCP peer review. And after a few seconds we are going to get a summary with the
[32:27] different KPIs for this code quality review. So you're going to see that Genie code presents the results really really nice using markdown structure. And at the beginning you can observe for instance the results from flick eight that we need to to correct some things
[32:45] that Genie code is analyzing these KPIs from the MCP. And also you can see that have some recommendations. And finally, as I said, we have a dashboard uh extremely dashboard on top of data bricks apps where we can analyze the
[33:01] different MCP results. So we can filter by by the project and we can analyze the KPIs over the time. So with idea of observe different patterns and different behaviors. And of course we have some detailed findings of
[33:18] this process. Okay, so finally next steps. So the first one is we want to deploy the full framework using continuous integration, continuous deployment. So the idea is we want to
[33:34] have version control for skills, prompts, and agent configuration. We can also run integration testing. So that means that we want to validate the agent outputs before any release. So the second one, we want to build a
[33:50] centralized and standard repository for skills. So that means that remember that we have a lot of libraries. So we want to establish a common pattern in order to build a skill or for our libraries.
[34:07] And of course control version of these libraries. The third one which is the most important for us, we want to run an agent evaluation. So that means that we want to build an agent just. That means that a large large language model as I just for validating the quality of the
[34:24] outputs. And of course human in the loop loop for the more critical decisions. And the last one, we want to move from pilot to scale. So, that means that we are performing onboarding of different business user teams. We also measure
[34:39] KPIs like time to model, good quality, and user satisfaction at higher scale. And finally, we want to scale based on we learn what we learn during this track. Um
[34:55] I think that that's all so from my side. So, I really hope you have enjoyed the talk and found it useful. Thank you so much and I think we have 4 minutes in case is there any question.

Learn more about the Databricks Data and AI platform.

The information provided herein is for general informational purposes only and may not reflect the most current product capabilities or configurations.