Transforming Financial Services in Latin America: Base39 Leaps to Efficiency with Amazon Bedrock

Initially, Base39’s clients had to perform loan analyses manually, limiting them to a maximum of 50 analyses per day per person. This process led to high personnel costs due to erratic staffing requirements—the volume of proposals varied greatly throughout the month—as well as delays in loan approvals of up to three days. Larger clients who invested in data scientists for traditional machine learning (ML) models still missed numerous loan approval opportunities due to the lengthy process of adjusting traditional ML models for specific rules—often taking months to implement due to extensive code deployment required for each new data source.

After consulting their technical teams about integrating Amazon Bedrock with MongoDB Atlas Vector Search, Base39 promptly decided to migrate, drawn by MongoDB's reputation for supporting innovation and security. The transition to a serverless architecture has proven smooth and flexible, allowing Base39 to utilize a comprehensive suite of AWS tools and services, including Amazon DynamoDB, Amazon Simple Storage Service (Amazon S3), Amazon API Gateway and AWS Step Functions. By relying entirely on AWS Lambda application deployment, Base39 ensures scalable and reliable performance that supports growth without degradation. This has been crucial for the integration of new technologies, allowing the organization to focus on delivering value while AWS supports the underlying infrastructure’s capacity.

Base 39’s broad catalog of APIs, empowered by Amazon API Gateway, enables seamless integration with existing applications or the use of a white-label interface, ensuring tailored solutions that offer high levels of customization and ease of use. Base39 uses Amazon API Gateway to connect directly to AWS services such as Amazon DynamoDB and Amazon Simple Queuing Service. By leveraging Amazon API Gateway’s integration with these services, Base39 eliminated additional compute layers, improving performance and cost efficiency while reducing latency. This direct connectivity enables faster data retrieval and processing while optimizing scalability and performance of the overall architecture.

For data management, Base39 employs Amazon DynamoDB for long-term caching, MongoDB Atlas as the primary database, and MongoDB Atlas Vector Search as the vector database. Secure backup, archiving, and the storage of important documents required for compliance with the Central Bank of Brazil are managed through Amazon S3. This configuration ensures both regulatory compliance and seamless integration with Amazon Bedrock, efficiently handling large data volumes and facilitating rapid access for AI-driven analyses. For credit analysis, Base39 quickly adopted Anthropic’s Claude 3.5 Sonnet and Claude 3 Haiku in Amazon Bedrock with support from LangGraph. This strategic combination enables Base39 to manage the complexity of loan analysis in a modular and adaptable manner, enhancing its ability to deliver precise and personalized credit solutions.

Providing ongoing guidance in the use of optimal tools and practices, the AWS Startups team was essential to their transition to Amazon Bedrock. The AWS commitment to innovation and provision of cutting-edge tools—like the workflow orchestration provided by AWS Step Functions—has enabled the Base39 team to better orchestrate AWS Lambda functions, creating more flexible processes and saving valuable development time. This optimization of their event-driven architecture has streamlined workflows and enhanced scalability, keeping Base39 at the forefront of the financial sector. Additionally, ongoing collaboration with MongoDB’s professional services team has been crucial in optimizing their architecture for the increasing volume of AI-driven requests, ensuring rapid response times.

The joint integration of Amazon Bedrock with MongoDB Atlas Vector Search played a pivotal role in enabling Base39’s rapid transition from zero to production in just two weeks, significantly reducing analysis costs and improving credit limits and approval rates through more contextualized analysis. In fact, Base39 required only one hour to seamlessly configure and test Amazon Bedrock with MongoDB Atlas Vector Search. This transformation has made Base39’s solutions more responsive, innovative, and customer centric. Streamlined workflows have allowed employees to deliver greater value and improve the overall customer experience with more personalized and efficient financial solutions. The company drives continuous optimization by quickly testing and implementing new models, with releases occurring almost weekly. They test at least three versions of models simultaneously each week to refine performance and optimize results, also exploring combinations with different Anthropic models.

With the implementation of Amazon Bedrock, Base39 has achieved a 96% reduction in cost per loan analysis. Previously, loan proposals could remain in the queue for up to three days; now, decisions are made in under an hour. Largely due to the rapid adaptability and scalability of Amazon Bedrock, the time required to implement new rules has been dramatically reduced from months to less than a week. Base39 also experienced significant reductions in infrastructure and development costs. Migration to a serverless architecture with AWS led to an 84% reduction in infrastructure costs and a 75% decrease in development costs for new features. Maintenance costs were reduced by 87-100%, allowing developers to focus on creating new features rather than maintaining legacy systems.

Reducing costs in their credit analysis process, Base39 effectively integrated two Anthropic models in their workflow: Claude 3.5 Sonnet for complex loan analyses and Claude 3 Haiku for quick responses to simpler questions at a significantly lower cost. Claude 3.5 Sonnet excels in complex reasoning and Retrieval-Augmented Generation (RAG), delivering responses 54% faster while maintaining costs similar to the original Claude 3 Sonnet model. Its ability to process large financial datasets and interpret nuanced regulatory and risk conditions is vital for accurate credit assessments, leading to informed lending decisions and reduced default rates. The flexibility of Anthropic’s model family is central to Base39’s strategy, enabling them to choose the most suitable model for each task in the credit analysis pipeline. By utilizing Claude 3.5 Sonnet for complex risk assessments and Claude 3 Haiku for simpler queries, Base39 optimizes resource use, lowers operational costs, and upholds high analytical standards.