My projects, from most recent to most formative.
Machine learning, deep learning, LLMs, computer vision and time series. Click a project for the full details.
Electricity demand forecasting for an island grid — EDF
Capstone project at EDF SEI: multi-horizon forecasting of electricity demand on the island of Ouessant, which is not connected to the mainland grid and runs mostly on diesel generators. Benchmark of statistical baselines, tree-based models (Random Forest, LightGBM, XGBoost tuned with Optuna), additive models (GAM with R/mgcv, TAM from EDF R&D) and a recurrent network (GRU), followed by online expert aggregation that matches or beats the best single model at every horizon. Code and data belong to EDF and are not public; the method and results are detailed in the technical paper.
- Online aggregation (MLPOL) of 5 experts: 3.32% MAPE at 30 min and 5.74% at 48 h over a 9-month test period
- Aggregation beats the best single model at 30 min, 1 h and 24 h and ties it at 48 h, with no need to pick a model upfront
Mini GPT — a language model built from scratch
Character-level GPT-style language model written from scratch in PyTorch without any Transformer library: tokenizer, causal attention, multi-head attention, Transformer blocks and autoregressive generation, each component hand-coded and validated with a measurement. Trained on Jules Verne's 34 French novels. Next steps: scaling to 2.7M parameters on Apple GPU, then studying the model's limits (knowledge probing, abstention, tool calling).
- Causal attention, multi-head attention and Transformer blocks written by hand, without nn.Transformer or any external library
- Test loss of 2.189 with 3 blocks, below the theoretical best-bigram ceiling (2.474) and far from chance (4.787)
Galen — AI assistant for medical diagnosis
Assistant for radiologists that analyzes medical images (MRI, CT scans and X-rays) with a multimodal LLM through the Anthropic API. Galen includes a conversational Q&A interface, generates PDF reports and ships a DenseNet121 model trained to detect pneumonia on chest X-rays.
- MRI, CT and X-ray analysis with Claude through the Anthropic API
- Automated generation of medical reports as PDF
Human action recognition in videos
Recognition of 101 human actions in videos with a 3D-CNN (R3D-18) and transfer learning from Kinetics-400, reaching 93.4% accuracy on the test set.
- 93.4% accuracy on the test set (1,723 videos) across 101 action classes
- Transfer learning from R3D-18 pre-trained on Kinetics-400 (400 classes, millions of videos)
Office attendance forecasting — Smart Workplace
Daily attendance forecasting in a smart workplace from weather, calendar and booking data. Comparison of classical and deep learning models, including a reproduction of the TimeXer architecture (NeurIPS 2024) on our own data. Joint project with Alexis Moisy (ISEN).
- XGBoost R²=0.888: best model across all approaches
- LSTM R²=0.528: best deep learning model, ahead of GRU, RNN and TimeXer
Flappy Bird — NEAT AI agent
AI agent that learns to play Flappy Bird with an evolutionary algorithm (NEAT) implemented from scratch in Python. No hard-coded rules: the agent learns purely from experience, generation after generation, by evolving a hand-written neural network.
- NEAT from scratch: neural network and genetic algorithm written without any ML library
- Architecture: 5 inputs -> 8 hidden neurons -> 1 output (flap / don't flap), tanh activation
Tech stack.
Tools and technologies I use for AI, machine learning and data science.
Machine Learning Specialization
Introduction to Deep Learning & Neural Networks with Keras
IBM AI Engineer Professional Certificate
Where I've built things.
My work in data science and applied AI.
Data Scientist - Work-study
- Electricity demand forecasting for island grids (Ouessant island), to schedule diesel generator start-ups
- Built predictive models for solar (PV) and hydroelectric production using weather data and tidal coefficients
- Designed API-driven pipelines that display equipment status and production in real time for monitoring
- Implemented energy optimization for smart appliances (heaters, EV chargers, water heaters, dishwashers)
- Data analysis and visualization for energy studies and diagnostics
- Automated recurring business processes
Interested in my profile? Let's talk about your project!
Get in touch →Let's start a conversation.
Looking for a Data Scientist role. I reply within 24 hours.
Available now for a Data Scientist role · anywhere in France or remote within Europe.