Machine Learning Engineer / Data Scientist — I build ML systems end-to-end: from raw data and feature engineering to trained, calibrated and deployed models. Physics & Astronomy background (Sapienza University of Rome).
| Project | What it is |
|---|---|
| ATP-Prediction-Model | Tennis match prediction: PyTorch MLP + custom Elo + isotonic calibration. V3 research track: 188 engineered features, LightGBM+NN ensembles, temporal splits, market-disagreement modeling — with honest metrics |
| chest-xray-classification | Multi-label chest X-ray classification (15 pathologies): Swin Transformer V2, asymmetric losses, Optuna HPO, GradCAM interpretability |
| mri-quality-assessment | Automatic MRI quality scoring (6 classes): focal-loss CNNs + Optuna, PyQt5 DICOM annotation GUI, exploratory SimCLR self-supervised track |
| tyre-lap-time-prediction | Regression on synthetic NASCAR data: paper-lite feature-package comparison — linear baselines win, and the report says why |
| NASCAR-AI-Strategy-Engine | Race strategy engine: XGBoost caution model, GAM tire models, Monte Carlo simulation, bootstrap decision analysis, live Streamlit dashboard (140 tests) |
| AdaptiveFraudAgents | Reply Code Challenge 2026: hybrid rules + multimodal LLM agents (GPT-4o investigation, Gemini audio transcription) with cost-aware design and LLM observability |
ML/DL PyTorch · PyTorch Lightning · scikit-learn · LightGBM · XGBoost · Optuna · timm Data pandas · NumPy · SQLAlchemy · ETL pipelines Product Streamlit · FastAPI · Flask · Docker Domain medical imaging (X-ray, MRI/DICOM) · sports analytics · tabular prediction
Methodology over leaderboards: proper temporal validation, leakage checks, probability calibration, class-imbalance handling, interpretability — and results reported as they are, including the modest ones.
- Email: leonardoschiavoni82@gmail.com
- LinkedIn: leonardo-schiavoni-665173340