Control & Automation Engineer turned ML/MLOps practitioner.
Focused on production-ready systems: pipelines that drift-detect, models that explain themselves, and demos that actually run.
🇧🇷 Based in Brazil · 🇮🇹 EU work authorization (Italian citizenship) · Open to remote
mlops-energy-forecast — End-to-End MLOps Pipeline
Automated retraining pipeline for energy consumption forecasting with drift detection.
Live: API docs · Dashboard UI
| Model | XGBoost · Train R²=0.989 · Test R²=0.673 (post-drift on purpose) |
| Drift detection | PSI + Jensen-Shannon divergence (threshold 0.20) |
| Orchestration | Airflow DAGs: daily ingest · weekly train · 6h drift check |
| Tracking | MLflow experiment tracking + champion/challenger model registry |
| Serving | FastAPI + Prometheus metrics + Grafana dashboard |
| Stack | Python · XGBoost · Airflow · MLflow · FastAPI · Docker Compose |
Airflow (schedule) ──► generate → validate → preprocess
│
◄── PSI / JS drift score ───┤
│ (every 6 h) │
▼ ▼
Trigger retrain Weekly retrain
│
▼
MLflow champion/challenger
(promote only if RMSE improves)
│
▼
FastAPI /predict /metrics
rag-sec-analyst — SEC 10-K RAG Analyzer
Ask natural-language questions about any public company's 10-K filing and get cited answers.
Live: API docs · Streamlit UI
| Embeddings | sentence-transformers/all-MiniLM-L6-v2 (384-dim, CPU) |
| Vector store | ChromaDB (persistent, cosine similarity) |
| Reranker | cross-encoder/ms-marco-MiniLM-L-6-v2 |
| LLM routing | Anthropic Claude Haiku → OpenAI GPT-4o-mini → Extractive (no key needed) |
| Evaluation | LLM-free RAGAS metrics via cosine similarity — runs in CI |
| Stack | Python · FastAPI · ChromaDB · Streamlit · Docker |
Evaluation scores (extractive provider, zero LLM cost):
| Metric | Score | Threshold |
|---|---|---|
| Faithfulness | 0.713 | 0.45 |
| Context Relevance | 0.437 | 0.35 |
| Answer Relevance | 0.585 | 0.40 |
SEC EDGAR API ──► chunker ──► ChromaDB
│
User query ──► embed ──► top-k*3 ─┤
▼
Reranker → top-6
│
Anthropic / OpenAI / Extractive
│
FastAPI + Streamlit
| Project | Description | Stack |
|---|---|---|
| ControlSystems-EN | Advanced Control Systems (Smith Predictor, LQR, PID) | MATLAB/Simulink |
| Predictive Fault Detection | CNN + Grad-CAM on thermal images — 91.2% ROC AUC | PyTorch · OPC UA · Jetson |
ML/MLOps: Python · XGBoost · scikit-learn · PyTorch · sentence-transformers
Pipelines: Airflow · MLflow · ChromaDB · FastAPI · Docker · GitHub Actions
Infra: Prometheus · Grafana · Render · Linux
Control: MATLAB/Simulink · OPC UA · MQTT · SCADA
Languages: Portuguese (native) · English (C1) · Italian (B1)
📬 lucaswilliamjunges@gmail.com · LinkedIn · EU work authorization — no sponsorship needed