School Projects

• Fine-tuning RoBERTa/Pythia
• Feature engineering
• Transformer embeddings
• Optuna
• Ensemble learning

Kaggle challenge (3rd year, École Polytechnique): classifying Twitter accounts as Influencers or Observers with an end-to-end ML pipeline — user-level feature engineering, RoBERTa/Pythia embeddings, Optuna-tuned LightGBM/XGBoost, and a bagged ensemble reaching 0.859 leaderboard accuracy.

• Time-series analysis
• GARCH
• Hawkes processes
• Quantitative finance
• Pair trading

Three time-series case studies (APM 52065, École Polytechnique): forecasting gold-futures volatility with a GARCH(1,1)-Student model, modelling crime contagion in Chicago with spatio-temporal Hawkes processes, and a copula-based VIX/RVX pair-trading strategy returning +93% over 2021–2025.

• Image segmentation
• U-Net
• YOLOv12
• Graph algorithms
• Link prediction
• Rule-based reasoning

École Polytechnique research project: an end-to-end pipeline turning handwritten sheet music into playable MusicXML — U-Net staff removal, YOLOv12 symbol detection, an MLP relation linker (0.86 Match+AUC on MUSCIMA++), and a rule-based assembler handling multi-staff grouping, polyphony, and tuplets.

• Generative AI
• Image segmentation
• Stable Diffusion
• LoRA
• Image inpainting

Generative modelling course (École Polytechnique): synthesising realistic masked faces from clean portraits with a two-stage pipeline — a U-Net mask predictor (IoU 0.93) followed by diffusion generation. A LoRA-fine-tuned Stable Diffusion inpainting model reached FID 17.7, beating DDPM-from-scratch and CycleGAN baselines.