Antoine Pizzetta

Certification project built solo. XGBoost model trained on five years of French property transactions (97 departments) enriched with OpenStreetMap proximity features, reaching R² 0.70 on a temporal validation split. FastAPI inference service, Streamlit dashboard, GitHub Actions CI with a blocking quality gate, drift monitoring (PSI, Kolmogorov-Smirnov) and automated retraining.

• Python/
• XGBoost/
• LightGBM/
• GeoPandas/
• OSMnx/
• FastAPI/
• Streamlit/
• Docker/
• GitHub Actions

Team project. A classification router sends each image to YOLOv8s detectors specialized by class family and fine-tuned on COCO 2017. Rare classes rebalanced with copy-paste augmentation, per-class AP evaluation, false positive and negative analysis, EigenCAM interpretability.

• Python/
• PyTorch/
• YOLOv8/
• torchvision/
• pycocotools/
• Grad-CAM

School project. A driving agent trained in the CARLA simulator through a custom Gymnasium environment, with a reward function built around traffic rules and varied traffic situations. Trained with Stable-Baselines3, learning curves tracked in TensorBoard.

• Python/
• CARLA/
• Gymnasium/
• Stable-Baselines3/
• OpenCV/
• TensorBoard

Audio transcription with Whisper, plus summarization and question answering over the recording, served by a FastAPI API with Redis caching. Azure virtual machine provisioned with Terraform, continuous deployment with GitHub Actions, load testing with Locust.

• Python/
• FastAPI/
• Redis/
• Docker/
• Terraform/
• Azure/
• GitHub Actions/
• Locust/
• Hugging Face

Client project at Eliophot. Retrieval-augmented chatbot answering guest questions from the hotel's website content, scraped with Puppeteer and indexed with LlamaIndex. FastAPI backend with response rules and guardrails for out-of-scope questions, Svelte and Astro front end, admin monitoring panel.

• Python/
• LlamaIndex/
• FastAPI/
• Puppeteer/
• Docker/
• Svelte

Team project on Spotify data from Kaggle. Compared regression models to predict track popularity (linear models, random forests, gradient boosting with XGBoost and LightGBM) and built K-Means recommendations, served through a Flask API and a React interface.

• Python/
• scikit-learn/
• XGBoost/
• LightGBM/
• Flask/
• React