Fintech for Doctors: Omnichannel via WhatsApp Using Amazon Bedrock

A Brazilian fintech built specifically for doctors. The platform combines technology and specialized services to simplify the operational side of a medical career, helping professionals structure and manage their professional activity, issue invoices, organize day-to-day financial flow, and stay current with tax and accounting obligations — all from an experience designed for the reality of doctors.

Beyond financial solutions, the company positions itself as a community for doctors: a space that connects professionals through shared challenges, knowledge exchange, and a support network that enables them to move forward with more agility and confidence. With strong execution pace and a product-led approach, it continues expanding its ecosystem to improve how doctors in Brazil handle the financial and administrative management of their careers.

Before this project, client onboarding happened through a web flow and relied heavily on manual processing. Users submitted documents and information online, and the team was responsible for validating data, confirming identity and professional registrations, and reviewing profile questionnaires. Over time, this model created friction at two critical points: user experience and internal operations.

From the user's perspective, the process was slower than it should be at a moment of high intent. When someone decides to sign up, any delay increases the risk of abandonment and frustration. Internally, manual validation raised operational costs and made onboarding volume difficult to scale. Additionally, collected data did not flow consistently to CRM workflows, generating rework, manual reconciliations, and inconsistencies between systems.

The goal was to bring onboarding to a channel where users already are in their daily lives — WhatsApp — while maintaining the rigor needed for document and business identity validation, and simultaneously creating a foundation that could be extended to other journeys throughout the client relationship.

Elevata designed and implemented an AI-based omnichannel onboarding capability, connecting the client's backend to WhatsApp and establishing a foundation ready to incorporate new channels over time. The aim was not simply to add a chatbot, but to build a complete onboarding system: capable of guiding the user step by step, validating information as it is provided, and triggering business rules reliably.

At the core of the solution is an architecture that processes WhatsApp messages through structured flows and executes onboarding logic via triggers. Conversation journeys are defined as flow files (in YAML), where the client describes what questions should be asked, what validations should occur, and how the flow should evolve based on user responses. Each message sent generates an event that can validate the response, call backend tools, and decide the next step in the journey. This ensures determinism where it is essential — data capture and validation — without compromising the naturalness of the user experience.

To enable channel integration and ensure future extensibility, Elevata used Amazon Connect as the channel layer, enabling WhatsApp initially and creating a clear path to support web and voice interfaces in the future without redesigning the core logic. Business rule execution was implemented with AWS Lambda, while runtime services were hosted on Amazon ECS. Natural language capabilities were provided by Amazon Bedrock, treating model selection as an engineering decision driven by latency and task complexity — rather than a one-size-fits-all choice.

Within Bedrock, Elevata implemented a two-model strategy to balance responsiveness and quality. Amazon Nova Micro was used for interactions that do not require complex reasoning, such as entity extraction, value normalization, and tool execution within the onboarding flow, with response times in the 500ms range. For steps requiring deeper reasoning, greater precision, and better language quality — such as summarization and information correction across a longer conversation — Claude 3.5 Sonnet was used, with typical responses around two seconds. This split maintained an agile user experience without sacrificing the accuracy needed at more critical moments in the process.

Equally important as the initial delivery was ensuring the solution could evolve. The flow-oriented model makes creating new journeys easy: simply define a new flow file, register the corresponding trigger, and test behavior before release. This way, the client can expand beyond onboarding without rebuilding the solution's foundation.