Selected projects
Described by the role I held rather than the code I wrote. Some I led teams through, some I built myself, roughly newest first.
5 selected works
- 01
AI adoption and governance across engineering teams
Designed and ran the adoption program for teams on critical projects, from the guardrails that define safe use to the training curriculum and the model for managing day-to-day usage. I measured the gains, and they showed a clear return.
AI governanceTeam enablementWorkflow designMeasurement - 02
Public health and education platforms at scale
Technical direction of the whole portfolio serving municipal and state services, reaching more than 2 million people a month. I set the architecture, the standards and the delivery priorities across the teams building it.
Ruby on RailsNodeReactPostgresAWS - 03
Agent platform with a built-in evaluation layer
In progressPlatform for corporate agents built on a single thesis: measurement is what makes cost optimization safe, because every saving, a smaller model, a cache, less context, is a bet that quality held. The agent loop runs end to end with streaming, tool calls and full tracing, durable runs on Temporal, and a deterministic provider that keeps it testable in CI without network.
Ruby on RailsReactTemporalAI APIs - 04
Natural-language analytics over the education data warehouse
Reads the transactional database, materializes flat tables in Parquet and answers questions in plain Portuguese about them, using text-to-SQL rather than RAG, because the questions are analytical and embeddings over rows miss counts and filters. Business rules live in SQL, not in the model, so the answer is the same every day and can be justified to a school, and it's now becoming a multi-tenant product.
PythonDuckDBParquetFastAPIAI APIs - 05
Educacenso pre-validation engine
Validates the school census file against the official INEP layout before submission, naming the exact field, the reason and the fix. The engine is deterministic, with rules extracted from the official spreadsheet into versioned caches, and AI only translates irregular rules into specs, once, without ever evaluating the file itself, which keeps every run reproducible and auditable.
PythonFlaskAI APIs
Earlier work
- Internal platform for AWS infrastructureGo, React, Postgres, Terraform and AWS
Infrastructure and applications provisioned through Terraform from a console instead of a shell, with encrypted credentials and an audit trail.
- Integration with Brazil's national health data network (RNDS)Ruby on Rails, Node and Postgres
The integration layer connecting public care systems to the national health data network, under its interoperability and compliance requirements.
- School enrollment portal for a state education systemRuby on Rails, Elasticsearch and Google Maps API
Includes the scoring engine that weighs questionnaire answers and geolocation to decide where each student is placed.
- Financial monitoring portal for a construction company.NET Core, React and Azure DevOps
Architecture and roadmap across backend, frontend and infrastructure, including the SAP and Salesforce integrations.
- Applied AI schoolReact, Hono, Cloudflare Workers, D1 and R2
A full learning product, catalog, checkout, student area and sales panel, with each lesson produced by an authoring pipeline, running end to end at effectively zero infrastructure cost.