AI product for the legal market

The product comes from a problem I saw up close in law and in my three years teaching lawyers electronic filing at the Rio de Janeiro Bar Association (OAB/RJ): legal work repeats a lot of things a machine does better, and the tools on the market were made by people who never filed a petition.

It is a RAG-based AI engine for law firms: case-law analysis, identification of arguments and risks, and drafting support. Multi-tenant, with privacy (Brazil's LGPD) built in from the start. It is where my two backgrounds meet.

Built as a pair with my business partner, Gilvan Borges, a software engineer focused on systems architecture (Java, hexagonal architecture, DDD). He designs the architecture and the back end; I build the AI layer: retrieval, evaluation and model safety. His work is at gilvanborges.com.br.

The split follows the architecture itself. In hexagonal architecture, the domain (the firm's rules, each client's data, what the LGPD requires) sits at the center and doesn't know a language model exists. The AI comes in through well-defined ports, like any other adapter. The idea is to be able to swap models, measure and harden the AI layer without touching the business rules.

Back end in Java with Spring Boot and front end in Angular, the stack where Gilvan and I are most productive. AI is part of the product, not a layer on top: language models with retrieval (RAG) over the firm's own material, so answers are anchored in the right documents.

In progress. This case study will grow with the architecture decisions, the results and what changes along the way as the product moves forward. The name comes out once the trademark is registered.