Machine Learning & Computer Vision engineer based in Paris. I take models from prototype to production and build what surrounds them: evaluation protocols that measure what a model is actually worth, and tools that let non-ML teams use it without me.
- 🧬 Bioinformatics Engineer (work-study) at Eligo Bioscience (2025–2026): instance segmentation in production (fine-tuned Cellpose-SAM), selective prediction with a reject option, Nextflow pipelines on AWS Batch.
- 🔬 Previously R&D intern at LIPADE (Université Paris Cité): multimodal Transformer fusing body pose, monocular depth and BERT embeddings.
- 🎓 MSc Computer Science, Vision & Machine Intelligence, Université Paris Cité · Erasmus exchange at Trinity College Dublin.
- 🌍 Open to ML / CV roles in Paris, the Alps region and French-speaking Switzerland.
| Project | What it shows |
|---|---|
| sits-crop-segmentation | Transferring video transformers (TimeSformer, V-JEPA 2.1 + LoRA) to crop segmentation on Sentinel-2 time series (PASTIS); a decoder redesign tripled the mIoU |
| simil.art | Content-based art image retrieval web app: ResNet features, Faiss indexing, refinement by style and color (Django) |
| ann-benchmark | Benchmark of 5 approximate nearest-neighbor libraries (Faiss, hnswlib, Voyager, Annoy) on 4 datasets: recall/QPS Pareto frontiers |
| stair-counting-rgbd | Classical RGB-D pipeline (CLAHE, Sobel, PCA, depth profiles) in C++/OpenCV, with a stratified evaluation |
| self-supervised-learning-comparison | SimCLR vs. rotation vs. inpainting vs. relative-position pretext tasks, plus CIFAR-10 → CIFAR-100 transfer |
| fuzzy-c-means-rust | Fuzzy C-Means from scratch in Rust for color image segmentation |
Methods: instance segmentation · object detection · self-supervised learning (TimeSformer, V-JEPA) · multimodal fusion · low-budget fine-tuning & domain adaptation · selective prediction & calibration · statistical evaluation (permutation tests, bootstrap)
🇫🇷 French (native) · 🇬🇧 English (fluent) · 🇪🇸 Spanish (intermediate)