Sponsored by Deloitte | Databricks for Banking ML: Accelerating Delivery with an AI-Enabled Platform

Summary

M&T Bank, a top-15 U.S. financial institution, is implementing a next-generation data science and ML platform on Databricks starting with 160 Treasury and Finance regulatory models previously running on a legacy on-premises Hadoop stack spread across more than 200 data marts.
Deloitte's AI-enabled 'Data Assist' asset uses an agentic ecosystem to reverse-engineer undocumented legacy code and data lineage, and forward-engineer modern Databricks implementations including Unity Catalog, MLflow MLOps, and a medallion architecture—with guardrails managing the non-deterministic behavior of LLMs in a regulated banking context.
An independent validation agent flags mapping and logic errors in reverse-engineered outputs, while an operating console monitors migration progress and quality metrics across all models in scope, establishing a single governed layer for modeling and reporting across the bank.

Chapters

FAQs

What is M&T Bank's goal for its Databricks ML platform migration?

M&T Bank is establishing a well-governed data science workbench on Databricks to serve as the foundation for its next-generation ML platform, supporting priority regulatory and growth use cases while replacing a legacy on-premises Hadoop stack. Phase 1 focuses on 160 Treasury and Finance models previously spread across more than 200 data marts. The overarching goal is to establish a single unified layer for modelers and reporting that enables MLOps across the bank for the first time.

How does Deloitte's Data Assist tool use AI to accelerate the migration?

Data Assist is an AI-enabled asset built on the Brickbuilder framework that uses an agentic ecosystem to reverse-engineer undocumented legacy code and data lineage from M&T Bank's existing Hadoop stack. It then forward-engineers modern Databricks code, including Unity Catalog definitions, medallion architecture transformations, and MLflow-compatible model implementations. The tool includes guardrails specifically designed to manage the non-deterministic behavior of LLMs in a regulated banking context.

How is data lineage handled during the migration?

This video demonstrates a reverse-engineered data lineage visualization generated by the Data Assist tool, which automatically traces how data flows through the legacy system's models and transformations. An independent validation agent reviews the reverse-engineered outputs and flags mapping or logic errors before they are carried into the new Databricks implementation. An operating console monitors migration progress, quality metrics, and validation results across all 160 models in scope.

What governance framework is being used for the migrated models?

The migrated models are governed using Unity Catalog for data and model discovery, lineage tracking, and access control, combined with MLflow for experiment tracking, the model registry, and promotion workflows. This establishes a medallion data architecture that provides M&T Bank with a single governance layer for both modelers and reporting users for the first time. The framework incorporates MLflow MLOps capabilities to support the full model lifecycle from development through production deployment.

Full transcript

[00:08] All right. Good afternoon everyone. How was lunch? Good. So, okay. Now I know who's going to fall asleep. That was a trick question. It's okay. Day three, lunch time. You can doze off through my section.
[00:21] Stay awake for these guys because he's going to have a some amazing demo on AI, right? So, my name is Fabio Italiano. I lead the uh data science platforms, which is
[00:33] Power BI, Tableau, Alteryx, Athlian, and now Databricks. And uh Shiram is my partner in crime. He's an MD at uh Deloitte Consulting. Long
[00:45] experience. I'm not doing any justice. Just look him up on LinkedIn. He's very impressive. So, this is about our story of implementing
[00:57] Databricks data science platform for my M&T Bank. M&T Bank is a super regional bank, mostly located on the East Coast. Um high concentration in Northeast. Our
[01:11] headquarters are in Buffalo. I spend my time between New York, Bridgeport, and Wilmington and Buffalo. That's where how my team is spread out.
[01:23] Uh we are in the top 15 uh in the US ecosystem. 20,000 plus employees, uh over $200 million in in asset. We grew mostly by
[01:37] acquisition. So, that's a an important element because that's uh what we had to contend with uh when we look at integrating all the systems. All right. So, today's talk is about
[01:52] going from data sprawl, uh silos, experience with data and modeling to bring in a every everything together into a unified layer of data and and
[02:06] machine learning platform. I This is just a snippet on one section of our data science uh uh, users. Uh, mostly the treasury and
[02:18] finance, 160 models, mostly on prem stack. Uh, we have over 200 uh, we used to have 200 data mart. We have a lot more data scientists that are
[02:31] not here. Our phase one will be focused on treasury and finance or regulatory models. All right, so this is basically our journey. Uh, so we use
[02:43] a Gentifi framework heavily to ingest data, to reverse engineer data, and also reverse engineer code from uh, legacy models.
[02:55] Our unified layer is we going towards a medallion architecture. That's uh the first time at the bank that we provide, you know, a single layer for
[03:07] modelers, for reporting, etc. And then last but not least, we're leveraging uh, all the capabilities of a best-in-class uh, MLflow
[03:20] uh, ML operations. All right, so Databricks enables all of this, all right? So, with the Unity Catalog, with the Gentifi framework, everything
[03:32] supports our vision. I'm not going to belabor this because as I said, we want to get to the demos quickly as possible. All right, and so in here
[03:44] I'm not going to explain what Databricks environment is, but for our quants, having an environment to they could do model development, model execution, and then have our risk to do model validation
[03:57] has been a struggle because they today they they do everything in separate environments. Now we're starting to bring in everything into one place.
[04:10] So with that said hopefully that was quick enough. Sriram is going to walk you through how we use Agent Tech on the data side and the modeling side. >> Sure. Thank you. Thanks, Fabio. Can you hear me okay?
[04:22] Okay. I want to make sure that all of us are still awake after lunch. So I won't bore you with more slides. I mean, we'll get to the demo as quickly as possible. So our vision what Fabio you know was
[04:36] said you know was saying in the beginning to set up a state-of-the-art data science platform at at M&T. Now some of How many of you are data scientists in this room?
[04:49] Okay. I've got a few. How many of data engineers in this room? Okay. So it's a good mix. So I'm sure all of you would have done you know gone through this. Primarily I mean your majority of time
[05:01] is spent in trying to get the data that you need than spending time in developing the actual models on top of it, right? I mean given a size of a bank like M&T, top 15 bank in the US and which is he
[05:13] made a point saying it has it has grown by acquisition, not organically. So you can imagine the they are different kind of data platforms that are brought in you know that So our foremost problem is how do
[05:26] we get data out of these platforms? And then serve it to the data science platform that you were trying to build. Now when we started this journey you know 6 months back we made a conscious decision
[05:38] we're not going to do this in a traditional data engineering way. How can we leverage agents to expedite this journey end-to-end, right? I mean given that we're building a data science platform where we're going to do a lot
[05:50] of machine learning and classic machine learning and deep learning and some of this genetic work, I mean, it'll be a, you know, atrocious to go and do this work and try and do the data engineering work in a in a manual way, right? So, right from the get-go,
[06:03] we set ourselves a vision saying, "Let's leverage agents for every single thing that we're trying to do. Let's try to automate this using agents as much as possible." So, that was our goal that we set. So,
[06:15] if you look at the two tracks, you know, I'll show you some of the demos as we go through. One track is for data engineering, which is the heavy lifting of the data sets that we need that are spread across the bank,
[06:27] on-prem infrastructure, you know, Amadeus is pretty, you know, old bank and you can think of like any company that has been in existence for more than a decade or two, you'd have like all legacy stack you could think of, right?
[06:40] Now, how do we use agents to, you know, understand this legacy data stack, reverse engineer all those data sets, forward engineer that into a modern, you know, cloud platform?
[06:52] Like, either you're trying to build our data products or or medallion architecture and building data products to the right side, can we leverage agents at every single step? That is one track of trying to reverse engineer the legacy stack and
[07:04] then forward engineer the legacy stack using agents. The second track is you can apply the same thing on the models. Today, these models are all, you know, before we start this journey, all these models have been developed using legacy
[07:16] stack on-prem, you know, you all be living, you know, living this, you know, journey for last decade. Can we use agents to understand these models? Because like any other bank, you know, people have people come and develop
[07:28] models, move on. Not enough documentation to clearly pinpoint, you know, what's really every single model is doing. There are documentations, but, you know, it doesn't always keep up with the pace at which, you know, you are changing the model.
[07:40] Can we leverage agents to understand these models, you know, document these models for us, forwarding this model to a Databricks PySpark, you know, notebooks, and deploy it on, you know, the Databricks platform data science platform that we're building. That's track B.
[07:52] So, right from the word get-go, we started doing this, okay, what are the agents we need to develop? What are the context those agents need to have? How do we make sure that these agents do not, you know, go off the rails and, you know,
[08:04] I'm sure all of you would have started developing your stuff with the agents. I mean, in your experience, in my experience at least, what I would say is the white coding thrill is always nice, right? You start developing something, you can white code it, you'll get an
[08:16] output in the week. You feel like, oh, you've achieved a lot, you feel like you've done a lot, and like once you start putting in the production stuff, you know, you start scaling it, that's when it starts biting you back really. You will start realizing, man, it's not actually doing what I'm,
[08:29] you know, what I'm asking you to do. You start seeing the gaps in the agents. We went through this journey for last 8 months, right? So, I'll talk you through some of those experiences, you know, what we did to make sure that the agents stay on the rails.
[08:41] How do we go from white coding to a production-grade, you know, code that you can really trust the output? What are the guardrails that we put in place so that we had the confidence that the output the agents are producing is actually reliable and actually
[08:54] repeatable, right? I'm I know this might sound counterintuitive, but the most important thing that we did when using these agents to develop a reverse engineer a foreign exchange code,
[09:06] how do I get a deterministic output from these agents? That was our goal because what we want to make sure is every single output these agents are producing are deterministic, repeatable, so that I can
[09:19] rely, you know, and it can produce a consistent result, no matter I keep running the same agents on the same set of same set of data sets, on the engineering side of the side, am I getting the same output? So, the goal was to produce a deterministic output,
[09:32] you know, from this non-deterministic LLMs and agents. So, that was one of the guardrails that we set to make sure that okay, let's have all the rails around these agents so that they behave in a deterministic way rather than producing outputs that I I can't really repeat
[09:45] from environment to environment or from run to run. Right? So, um let me move on. So, let's talk about the engineering agents, right? So, as we go through the demo, I'll try to rush through the slides so that we
[09:57] have more time for the demos. So, this was a pretty complex exactly any other firm that you look at ecosystem. You name the stack, a SQL-based stack or a, you know, a
[10:10] Hadoop-based stack or, you know, this the legacy data stack was all over the place. So, our first job was like come up with the ecosystem where these agents can work with each other. So, you have
[10:22] set of agents that can go and reverse engineer the code base, produce a consistent output and a reliable output, a deterministic output that gives us the confidence that yes, we understand the legacy stack that is sitting there. We
[10:34] understand the data set that's being produced. So, that we can and the documentation is clear enough for us to, you know, pinpoint every single attribute of data that's coming through so that my data scientists when they start using the data sets, if
[10:47] there is a gap, we can exactly pinpoint where the gap is coming from. So, that is the first set of agents we developed in reverse engineering the existing data engineering stack that was running on prem. The second set of agents is like, since, you know, as Fabio mentioned, they grew
[11:01] both organically and mostly through acquisition, when you grow through acquisition, you're going to have duplicate data sets sitting across across the bank, right? You're going to have the same data set coming from, you know, X system over a
[11:13] period of time someone went and developed in Y system. You're going to have overlaps. You never had time to go and, you know, fix those overlaps. So, the code base was also like was, you know, such a complex spaghetti that now
[11:25] you need to understand and consolidate and get a clear picture of how what data sets are coming through. So, that's a second set of consolidation set of agents that we developed to once a code is reverse engineered, now let me consolidate the code base to clearly
[11:37] understand what's going on in the code base. The third set of agents that we developed is basically forward engineering agents. Now, you develop all this, you know, you got all this information, how do I modernize this code base? Like can
[11:49] these agents now build it in a medallion way, you know, in a cloud platform? Can it produce data products that I need rather than if you look at the legacy code base, you might have like 10 hops, 11 hops, 12
[12:01] hops, 15 hops as data goes through, you know, because people have developed it over a period of time. Can we harness all this intelligence? Can I put context around all this? Can my agent now develop a much more modern code base that can run on a medallion
[12:13] architecture? That was the third set of agents that we developed. So, you can apply the similar concept for the model code base, right? I mean, once one side of the track was the engineering code base, all the data engineering pipelines, you know, that
[12:26] has been running. We did something similar on the model code base, which was I mean, if you have been developing models in legacy stack, I'm sure you will have a mix of ETL ELT code running in the same
[12:38] platform. You'd have data science model code running in the same platform, and you'll have a mix of these scripts, you know, it's very hard to differentiate, you know, okay, which script is doing engineering data engineering, which script is doing my data science. There
[12:50] are some methods and stuff, but you start realizing like, you know, it's been developed over a period of couple of decades. How do I isolate the workload? How do I isolate my data science workload versus my data engineering workload? So, that they belong, you know, that way I can
[13:02] isolate my data science workload, put it on a MLops pipeline on Databricks that run within that space, and I can, you know, differentiate my persona of my data scientist focusing on all the data science modeling versus my engineers focusing on the engineering
[13:15] aspects of it. So, we developed similar set of agents that can dissect my model code, produce the same reverse engineering output, differentiate that into engineering code and sign a data model code, a data, you know, modeling
[13:27] code. And then forward engineer that into a modern Databricks platform. So, that's a second ecosystem that we developed, you know, to do all this from end to end.
[13:39] So, without further talking, maybe I'll just quickly show some of the stuff.
[13:56] Okay. This is think of like a backbone that we used across end to end, right? >>
[14:09] can do deterministically what you want it to do?" So, we developed this toolkit. I mean, this is an agentic toolkit that has set of agents, a set of skills, pipelines, templates, context, and
[14:22] knowledge base, and they all share this ecosystem. So, that way, what it helped and is still helping us, I mean, we're still in the journey, you know, um and we are doing more on the same ecosystem to develop more of agentic
[14:34] stuff. This ecosystem that is sitting in the back, you know, on the back of it, is where we're developing all our agents. That way, an agent that is doing reverse engineering on a X stack can share the same context with another agent that is
[14:46] doing reverse engineering on Y stack, and you have those lessons learned, human in the loop information that was put in. So, these agents now have a common way of talking to each other. I mean, I'm not going into MCP and all of
[14:58] that stuff, but think of like you have a a common ecosystem where these agents share the context. They have skills they have developed over a period of time. They can share the information from one agent to other agent. I can develop more
[15:10] agents rapidly based on the tasks that I'm seeing on the ground, right? And I'll show you some of these agents, you know, as we as we go through the demo, right? This was our backbone, you know, that on which we started building the entire platform where we're doing reverse engineering, and forward
[15:22] engineering of the data engineering pipelines, and similarly, reverse engineering and forward engineering of the data model modeling code that we're developing.
[15:35] So, I'll take one example. For example, let's just say So, we look at this this whole thing that I'm showing on the screen was the output of my Now, the reverse engineering agents that went and reverse engineered the entire stack, identified
[15:48] the domains, you know, lending facilities, parties, you know, loans, etc. All this output was produced by the set of agents that went and reverse engineered the entire code base, and the lineage that's being produced end to
[16:00] end, you know, um how this particular data set is being produced, what's the input, what's the output. So, this entire ecosystem that you see here
[16:12] was basically those set of agents. And if you can see, there are like 900 plus table, 1,300 connections, 17 domains. So, the agents have gone through each one of those things, understood each of the data sets that are being produced, how the data set is being, you know,
[16:25] processed from one intermediate hop to other intermediate hop. I mean, if you are I mean, if you're developing modeling code, if you're developing scripts, I'm sure you create those intermediate data sets that feed one You create one data
[16:37] set. That data set feeds another data set. I'm sure you're creating those macro data sets and all of that stuff, right? All of us are very familiar with this. That's exactly what we were going through. Analyze the entire ecosystem, understand how each of those, you know,
[16:50] data sets are being, you know, leveraged. What's the input? What's the output? What's the logic? You know, all of that stuff. So, this Imagine doing this in the traditional manual way of, you know, analyzing every Of course, we all
[17:03] have been doing automation, but the advantage that these agents have started giving now is like, you know, they have their reasoning capabilities have improved a lot in the last like 18 months or so, right? I mean, if you ask me the same question like maybe 18 months before, we were not getting this
[17:16] level of intelligence with these agents, the reasoning capabilities and correcting their own mis- mistakes, finding the gaps in their own analysis have improved dramatically in the last 18 months. So, what was not possible 18 months before, now, in terms of the reasoning
[17:30] capability these agents have, in the trade to produce this accurate analysis, has improved a lot. I mean, you What we have been using is a combination of, you know, Claude and Codex, but, I mean, pretty much you you can name the
[17:43] and now we are putting all this into the as we move forward. Now, I think you might have sound sound of you have not seen announcement last yesterday, where the Omni agent is now having a harness on top of all these agents, right? So, we have an experiment with
[17:55] that, maybe, that we will do next. But, currently, this is all running on individual agents, you know, being a combination of Claude and Codex, doing all the heavy lifting for us.
[18:09] So, now you're reverse engineering it. Let me just go even further. Now, all the visualization, the lineage, everything is being produced by the agents. See, none of this stuff is what I'm showing you on the screen has been hand coded by any of the engineers. Of
[18:22] course, we worked hand in hand with the agent, we fixed a lot of that stuff, but the heavy lifting heavy heavy lifting has been done by the agents, right? Now, if you look at these, you know, um individual tiles here,
[18:34] these are all the individual systems of records that we analyzed as we are building our engineering pipelines and modeling pipelines, right? I mean, each one of those tile stands for an individual system of record from where we have got getting our data sets,
[18:46] analyzing those pipelines end to end, understand what's going on those pipelines and forward engineering those pipelines, right? So, if I take, for example, let me just go a random example, maybe RCC.
[19:05] all the visualization that you see here, this is again, it is reverse engineered the entire pipeline, telling me how the data is moving through the legacy stack pipes. This gives me if you're an engineer
[19:17] you're given a large complex code base and you don't have much of knowledge and information the more visual cue the agents can give you to understand the legacy system. That's what these agents have developed. Now, if you look at the Sankey visualization, it tells me that
[19:29] majority of data is coming from these source systems, you know, it's going through these pipes and that majority loading into my account, you know, data set that I'm creating here, right? Now I'll go to the forward engineering part of it now. So, the left-hand side is
[19:42] what you see is the legacy stack, what the agents have reverse engineered. The right-hand side is what they have forward engineered that into a a medallion architecture, right? So, if you see if I zoom in further on the right side
[19:55] I'm sorry, I don't know why this is not clicking. Let me give you a second. Okay. You see the bronze layer. You see the silver cleanse layer. And you see the final silver confirm layer. So, the legacy stack what you see 1 2 3 4 5 6 7 or 8 hops on the left side
[20:09] was completely reverse engineered, documented at the attribute level. If I go here for a second every single field has been documented how, you know how are we moving that
[20:22] into the target state? Are we refactoring it? Are we matching it? And are we removing it? And then this gives me the lineage flow how this spaghetti flow has been re-engineered into a a medallion architecture, you know, on the right side.
[20:34] And like we even developed an an entire ecosystem of how do we validate the outputs of these agents independently than the work done by these agents, right? So, think of like
[20:47] you have set of agents. They are going and reverse engineering a large complex code base passing out that information to another ecosystem of agents that can now forward engineer that code base to a modern platform.
[20:59] We developed another independent set of agents. Can it go and verify every single step that these agents are doing and identify issues and provide feedback and fix that independently rather than you know, hallucinating along with these
[21:11] agents. Right? So, we developed an independent you know, system that think of like a set of agents whose sole purpose is to go and verify every single thing these what these
[21:24] other agents are doing. So, you go and verify is your field lineage complete what you're claiming it to be? Is your transformation accurate what you are trans what you are translating from
[21:38] your legacy stack to the feature stack? Right? Is your is is your code coverage complete in terms of what you're reverse engineered and what you're forward engineering? Are there gaps in this translation from
[21:50] legacy to the modern modern place? So, these agents' job is to question every single thing what the other agents are doing. Come up with the metrics that we can accurately measure and feel confident that yes, this is doing what it's
[22:03] supposed to do. I mean, were there issues? Absolutely. I mean, were there limitations in what these agents agents can do? 100%. But, did it give us a lift than doing manual effort? 100%.
[22:16] So, some of the metrics that we developed is to verify do I have I'll show you some of these metrics uh shortly. Do I have leak freedom? Now, what I mean by that is is my legacy code leaking
[22:28] into the forward engineering? You will you'll see a lot of these observations as you go through. These agents are very confident in claiming what they do. Right? Till you start questing them, they'll start agreeing with you. Right? And I mean, you would have seen this behavior again and again. You start questing them
[22:41] saying, "Hey, you know, you have taken a Goovi based logic, you know, that is something that is specific to that tool. Let just say you're saying you're writing a in a simple sequel, you're writing a case when statement. But, in the Goovi based tool, you will drag and drop something to do the same
[22:53] case when statement. These agents can then take that Goovi based logic and then put that as a sequel logic and claim that they have done their job. So, those are the some of the issues that we ran into. We start questioning, we're developing independent metrics
[23:06] that can question these things and give us the confidence that, fine, the system is working as expected.
[23:22] So, this is kind of think of like an operating console that we developed. This is the independent agent ecosystem that was developed to verify all the coding the Z other agents are doing. The green boxes are the entities and SORs that are not issues. The yellow ones that are now have issues. You see the
[23:35] operating queue on the right side. I mean, that is currently empty. But, this is where humans and agents come together. So, we have a team of engineers who look at these dashboards, look at their operation queues to see which one of those entities and code bases that being
[23:48] translated is being modernized have issues. What are those issues? And they start clearing those queues, providing the feedback back to the agent either by providing more context or providing more prompts, you know, tweaking some of the underlying architecture behind these agents to come
[24:01] up with a much cleaner output what what we have been observing. Right? So, maybe I'll show some more examples if I can go there. So, for example, if you see here on the top,
[24:14] you have 377 entities. It now this is a snapshot that we have taken in a few weeks back. But, you look at the dashboard, there are 377 entities we're trying to reverse engineer and forward engineer. There are 331 that are healthy.
[24:26] 46 needs reviews. There are some advisory signals the agents are giving, "Hey, you better go check it out." May not be an issue, but I would appreciate if an engineer, you know, human engineer goes and takes a look at it and then say, "This is actually, you know, doing something."
[24:39] If you look at some of these errors, for example, if I take party account as an example, you could clearly see it is trying to reverse engineer and forward engineer a legacy UI-based ETL tool. It is trying to
[24:52] forward engineer that into a PySpark in the future. And it is not able to bind some of these parameters and it throw an error back to the engineer saying, "Hey, I don't know what this dollar dollar param capacity code filter means. I want a human to go
[25:04] take a look at it and give me the information before I can translate it further." Right? So, there are several such issues, you know, um And these are the metric systems that I was talking about like the translation correctness,
[25:17] the graph completeness, the leak freedom, you know, there are a lot of metrics that we developed over a period of time which are very specific to the problem we are trying to address, right? So, what I'm trying to say is
[25:29] in the beginning I showed you the entire agentic ecosystem. There are a set of agents, skills, templates, contacts, all of that stuff. We are trying to address a very specific engineering problem of how do I take a legacy code base, modernize it on, you
[25:41] know, cloud platform, develop my models on Databricks and put it on an ML platform. While we are doing this, we understood, you know, okay, to make sure that these agents do not hallucinate, these agents do not give us a confidently false answer saying that
[25:54] yes, I've done it. Our second time you question it, they start agreeing with you. Develop an independent metric system that gives us the confidence that, you know, this whole thing is working as expected. We just need to go do some manual intervention where needed to give
[26:06] us the lift that we need. So, these are the metrics that we developed which is independent by these agents we go and develop. I'll show you some of these agents, you know, at the high level. How much time I have? Okay.
[26:19] So, if you look at these some of the agents on the left side, I'm just showing you some of the empty files. I mean, if I go to this agent called Sentinel, for example. Is it Is it readable on the screen? I hope it is. Um If you look at these agents'
[26:31] personality, we call it as you're a skeptic. You assume every mapping is wrong until proven right. You assume every data convention will lose precision. You assume every function translation will have edge cases that will fail. So, this
[26:43] is a core principle of this, you know, we have set for these testing agents to go and question everything the other agents are doing. And some of these The agents that you see here, let's go here, Spark, for example. This agent's job is to live and breathe
[26:55] Databricks. Right? I mean, you know the Delta Lake internals, you know Unity Catalog governance, you understand the ETL pipelines. So, these agents job is to make sure that the forward engineered code that we are producing is optimized for to run on Databricks.
[27:09] So, you know, we have 12 minutes left. Maybe I'll uh Just one second. I'll just hand it over to Fabio and then we'll take questions. Uh I know we only showed you the glimpse of
[27:21] what the idea is to show the art of possible what we are doing. Not like we won't be able to get you know, get down to every single detail out there, but hope you know, you you get the idea what these we are doing with this and what we can do with this.
[27:34] I'll hand over to Fabio. Do you want to take it from here? >> Uh just uh one word on the metrics. Uh
[27:47] so, our bank is very old and very conservative with our leaders. Uh we follow you know, we're very risk-averse. We manage our everything
[28:00] uh with an eye on uh on risk, right? So, even when we put out these measures, our leadership was like, "Are we sure these are real?" And they are real, right? But maybe by a approximation
[28:12] uh especially compared to the journey that we started 3 years ago on building our medallion, this is a significant a significant step uh forward. Uh in terms
[28:25] of speed, quality, and hopefully the durability in the future. So, want to I'll leave this on and uh
[28:37] maybe we can open for questions.

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.