Portfolio · LinkedIn · Email · Open-source contributions
I'm João Gabriel, a software engineer based in Brazil, building professional software since 2021. My foundation is PHP, Laravel, WordPress/WooCommerce, and relational databases; my independent projects extend that experience into Python APIs, data pipelines, and source-grounded LLM applications.
I focus on the parts that make a system useful beyond a demo: clear data contracts, background jobs, failure handling, tests, and evidence users can inspect.
From changing public-exam notices to an app that works offline.
A monitoring platform for Brazilian public-exam announcements, with a catalog of 99 configured sources and a React Native companion app.
- Data pipeline: HTTP, browser, and PDF collectors feed a FastAPI backend. SHA-256 fingerprints and version records track document changes; crawl diagnostics distinguish empty results from blocked or failed sources.
- Offline workflow: the app reads from SQLite, queues favorites and alerts in a local outbox, and synchronizes them with the backend when connectivity returns.
- Inspect the implementation: source catalog and diagnostics, mobile sync design and limitations, and roundtrip verification scripts.
Python · FastAPI · PostgreSQL · Celery/Redis · React Native · SQLite
Compare RAG and Full Context, then follow an answer back to its evidence.
An interactive research archive using Gemini, pgvector, and a Next.js interface. Each run records its selected sources, citations, latency, token usage, and estimated cost.
- Evidence: the backend resolves citation identifiers to exact source passages and positions. A second model call reviews factual claims against those passages before publication.
- A measured constraint: the 11 October 2026 verification indexed 738 chunks from a private three-book corpus. Whole-book Full Context attempts hit HTTP 429/503 responses, leading to a 30,000-token input guard and explicit scope reduction.
- Reproducibility: the report connects real provider responses to resumable indexing, citation corrections, and documented failure modes. Live runs and simulated demonstrations are labeled separately, with the conditions and limits of the comparison available to inspect.
Python · FastAPI · PostgreSQL/pgvector · Gemini API · Next.js · TypeScript
- GymIA: an academic workout-planning prototype for physical education professionals. OpenAI generates structured plans that instructors can review and edit, alongside client management and MySQL persistence. React · Node.js/Express · Prisma
- Android Fitness Ecosystem: Evolução, NutriTrack, and Coach connect workout logs, nutrition, Wear OS, and Health Connect. Local storage supports core workflows; AI features call the Google Gemini API using the user's key. Java/Kotlin · WebView · SQLite
| Project | Contribution | Pull request |
|---|---|---|
| jvm-mcp | Probe Spring Actuator before opening JMX, with fallback and resource-cleanup regression tests. | #85 |
| AWS GenAI DB Modernizer sample | Extend AST checks and mutation fixtures to detect inline database-engine rules. | #1116 |
| Repowise | Normalize the dead-code scoring penalty by symbol count, including empty-module and saturation cases. | #4083 |
| Vavr | Test null error-mapper results in Validation.mapError and bimap. |
#3372 |
Backend: PHP, Laravel, Python, FastAPI, REST APIs, WordPress/WooCommerce
Data & operations: PostgreSQL, MySQL, Redis, Celery, Docker, GitHub Actions
Applied AI & interfaces: RAG, LLM evaluation, pgvector, TypeScript, React/Next.js, React Native
Academic background in Information Systems. Interested in backend and applied-AI work where data quality, traceability, and maintainability matter.