AI Engineer building LLM agents, RAG systems and AI guardrails & evaluation. I work on the part after the model works: grounding, verification, and the limits that keep it honest.
Open to remote roles, international or Brazil-based · GMT-3
| What it is | Proof | Stack | |
|---|---|---|---|
| sb100_agents | Self-hostable RAG API that scores its own confidence | Agent gate calibrated to 96.7% / 3.3% TPR-FPR · 370 tests, 90% line coverage · 15 ADRs | Python FastAPI Qdrant Ollama LangGraph |
| tweet-sentiment | Sentiment classification fine-tuned end to end | 0.887 macro F1, +51.9% over the frozen-features baseline · 4.1x faster Rust preprocessing at 1M tweets | Python Rust PyTorch Polars |
| weather-forecast | scikit-learn model running in the browser | Matched to 1e-6 against Python · 200-tree forest serialized and traversed in TypeScript · test-first across 20+ features, traceable in the log | Python TypeScript LightGBM |
| visiosoil-app | Soil texture from a photo, fully offline | 3rd of 1,300+ at FETEPS 2025 · poster at ICPA/ConBAP 2026 · 26 on-device descriptors, resampling byte-identical to Pillow · 992 Dart + 810 Python tests | Flutter Dart scikit-learn |
| my-framework | Engineering standards, enforced inside AI agents | v0.8.0 · 7 quality gates · 21 Go test packages | Bash Node.js GitHub Actions |
AutoMQ #3261 |
Accepted contribution to a Kafka-compatible streaming platform | Response-size validation so a request/response mismatch fails cleanly instead of crashing, shipped with 157 lines of tests | Java |
Closed source. Numbers come from my own commits in those repositories.
| What it is | Proof | Stack | |
|---|---|---|---|
| Agent code-generation platform | Idea in, verified application out | Replaced an LLM-driven loop with a deterministic 4-wave scheduler (up to 12 agent phases; 12/14 E2E runs approved) · security gate blocked 21/21 severe findings in a 22-project pilot · 9-layer guardrails · auto-repair fixed 14/21 gate failures | Python LangGraph Semgrep Trivy Cloud Run |
| AI college-admissions advisor (early-stage U.S. edtech startup) | Grounded answers over ~1,800 institutions | Hybrid retrieval over pgvector · refuse-before-call gate and post-answer verifier · CI footprint ~9x smaller · 561 tests at launch | Python FastAPI pgvector Anthropic |
Agents & Orchestration
RAG & Retrieval
ML & Inference
Backend & Data
Cloud Data
Infra & Quality
