Itaú Improves Speed to Market and Productivity of ML Solutions Using Amazon Web Services

Itaú provides banking services to customers in Brazil, Latin America, and 18 other countries around the world. It has over 95,700 employees, about 15,000 of whom work in IT. Itaú’s original infrastructure was entirely on premises, which led to high costs and slow development times. The on-premises infrastructure was also not scalable because it was limited by physical space and hardware. In the bank’s data group, data scientists would have to wait up to 6 months for memory and resources to be made available, and the company had a deployment waiting list with over 100 ML models on it.

To overcome these challenges, Itaú decided to migrate a portion of its business to the cloud and chose to use AWS. “One of the reasons we chose to migrate from on premises to the cloud was a strategy to increase business competitiveness and efficiency at the same time,” says Diego Nogare, ML engineering manager at Itaú.

Around 6 months after the start of the migration, Itaú chose Amazon SageMaker—a service to build, train, and deploy ML models for any use case with fully managed infrastructure, tools, and workflows—as a flexible, cloud-native ML solution. “We were transforming our software and data using AWS, and we needed a solution that worked perfectly on AWS,” says Vitor Azeka, data science superintendent at Itaú. “Amazon SageMaker was the obvious choice.” As of 2024, around 60 percent of the company’s software and data is already modernized to run on cloud”.

Itaú has built a complete solution for its data scientists using AWS. First, data is gathered using AWS Glue, a serverless data integration service that makes it easier to discover, prepare, move, and integrate data from multiple sources for analytics, ML, and application development. This data is then used to start experiments using Amazon SageMaker Studio. Itaú uses Amazon SageMaker Studio as a flexible development solution for its internal data scientists to experiment. Next, ML models are deployed using other Amazon SageMaker tools, such as Endpoints, Batch Transform, and Asynchronous Inference. The company monitors models using Amazon CloudWatch, which collects and visualizes near-real-time logs, metrics, and event data in automated dashboards to streamline infrastructure and application maintenance. Using all these AWS services together, data scientists can achieve their needs.

Itaú delivered its first solution using Amazon SageMaker Studio as an integrated development environment in August 2021, and as of April 2023, it had more than 3,200 unique users for the AWS service, including around 350 data scientists.

Itaú no longer has a waiting list for deploying ML models. Using Amazon SageMaker Studio, the company has reduced deployment time from up to 6 months to 3–5 days in some cases. This reduced deployment time improves speed to market for the company. “When we use Amazon SageMaker Studio, we can run our pipeline and deliver the solution to our customers very quickly,” says Nogare. “We can thus improve customer experience.” Itaú is also saving on costs compared with its old on-premises infrastructure.

Since November 2021, Itaú has had weekly meetings with the AWS team to discuss architecture, security, and its road map. “AWS support was very important to achieve the results we have today,” says Nogare. “Any time we saw a problem with our solution or governance needs, the AWS team supported us through that.” Some of the governance needs are met using Amazon SageMaker Studio. When the company runs pipelines to provide Amazon SageMaker Studio to users, issues of governance and security are already resolved.

The standardization of its solution means Itaú can more easily onboard new employees and move data scientists from one department to another. Updating is easier since everything is virtual, and the company no longer needs to rely on physical machines. Using AWS, the pipelines for data scientists are integrated, so ML models are deployed and monitored in the same data pipeline. This further improves efficiency for data scientists.

“At the end of the day, we can deliver faster,” says Rodrigo Fernandes Mello, distinguished data scientist at Itaú. “We have improved standardization and integration, and we can use AWS to continue improving.”

Itaú is looking to continue improving its standardization. The next step for internal standardization for its data scientists involves having more employees using its IARA solution, which is based on AWS and uses multiple services, including Amazon SageMaker Studio. Itaú will continue to develop its pipeline by using tools inside Amazon SageMaker, such as Amazon SageMaker Pipelines, which is used to create, automate, and manage ML workflows at scale. Itaú is running tests to bring more standardization to its pipeline using AWS tools.

“This project brought a lot of efficiency to the data scientist team,” says Azeka. “Using Amazon SageMaker Studio, we can test new things while publishing others, and we can discuss state-of-the-art solutions using large language models. This makes our data scientists proud to work at Itaú.”