Projects

Open source, with tests, docs and deploys. Each one tackles a problem I see in production.

vicgamelo/agent-orchestrator

Multi-agent API: a coordinator delegates to specialists in parallel, with MCP tools, guardrails and end-to-end tracing.

  • Supervisor/worker delegation with parallel tool calls
  • Tools from any MCP server, namespaced per server
  • SSE streaming, and every run becomes a trace in genai-lens
  • Runs on Claude, OpenAI-compatible APIs or local models
  • Agents
  • MCP
  • FastAPI

vicgamelo/mcp-server-brasil

MCP server with Brazilian public data for agents: CEP, CNPJ (alphanumeric too), holidays, business days, Banco Central and IBGE.

  • 8 tools, a resource and a prompt
  • Validates the alphanumeric CNPJ that starts in 2026
  • Caching, retries and an HTTP mode with per-IP rate limiting
  • MCP
  • Open source
  • Python

vicgamelo/agent-auth

Authorization for AI agents and system-to-system calls: a small OAuth 2.1 server following the MCP spec, a guard for MCP servers and APIs, and token exchange that records which agent acted for which user.

  • PKCE, one audience per resource and 5-minute tokens
  • Token exchange (RFC 8693) with an actor chain, never forwarding the token it received
  • Revocable consent that cuts access within seconds
  • Hash-chained audit log
  • Attack tests: replay, PKCE, audience, confused deputy, SSRF and registration floods
  • OAuth 2.1
  • MCP
  • Security

vicgamelo/rag-normas

RAG over the Banco Central do Brasil's Pix regulations, built only from open-source parts: hybrid search in one SQLite file, ONNX models on CPU, the act and article cited in every answer, and an evaluation that runs in CI.

  • Indexes only the text in force: the superseded wording the BCB keeps struck through stays out
  • BM25 + vectors + re-ranker: the right first result goes from 60% to 80%
  • Every sentence cites [n]; invalid citations are flagged and out-of-scope questions get "NÃO ENCONTREI"
  • An MCP tool for agents, traced in genai-lens
  • RAG
  • Evaluation
  • MCP

vicgamelo/genai-lens

Self-hosted LLM observability: traces, tokens, cost and latency per model, with an OpenTelemetry GenAI endpoint and PII redaction.

  • Python SDK that never takes down the app it's observing
  • Ingests OTLP/HTTP using the OpenTelemetry GenAI conventions
  • Dashboard with trace waterfall, cost per model and time series
  • Observability
  • OpenTelemetry
  • LLM

vicgamelo/mlops-platform

End-to-end MLflow for credit risk: traceable training, champion/challenger promotion gated on quality and fairness, drift and serving that swaps models without downtime.

  • AUC 0.77 on German Credit, with a 0.004 gap between age groups
  • A model is only promoted if it beats the champion without hurting calibration or fairness
  • One-command rollback and PSI drift detection
  • MLflow
  • MLOps
  • scikit-learn

vicgamelo/ocr-brasil

In-browser OCR for Brazilian documents: finds CPF, CNPJ and boleto lines in a photo and uses the check digit to decide what to trust.

  • Tesseract (WASM) with the Portuguese model served by the site itself; the image never leaves the device
  • Fixes O→0, l→1, S→5 only when the check digit agrees, and re-reads the strip that failed
  • Adaptive threshold (Bradley-Roth) for photos with shadows
  • OCR
  • Computer vision
  • TypeScript

vicgamelo/giz

Visual AutoML that runs in the browser: from a spreadsheet to a trained model, with Python in WebAssembly and a local AI assistant. The file never leaves the device.

  • Data profile, cleaning recipe, AutoML with a live leaderboard, evaluation and batch prediction
  • pandas, scikit-learn, LightGBM and XGBoost on Pyodide; works offline once loaded
  • An assistant that runs on the GPU (WebGPU), explains alerts and metrics, and only acts on a click
  • Exports the Python script that rebuilds the model and logs the experiment to MLflow
  • AutoML
  • WebAssembly
  • WebGPU

vicgamelo/ml-from-scratch

Classic ML algorithms and attention in pure NumPy, each one checked against scikit-learn.

  • Decision tree and gradient boosting matching scikit-learn exactly
  • K-means++, PCA, metrics (AUC, ECE) and a Transformer block
  • NumPy
  • Fundamentals