Final-year engineering student at EPITA (IMAGE track), in the M2 MIND programme at Sorbonne University, looking for a PhD in medical imaging starting in 2027. As a student researcher at the EPITA Research Laboratory (LRE), I currently work on learning hierarchical structures with recurrent graph neural networks.
Email · LinkedIn · RFIAP 2026 paper (HAL)
- Lεtree (Learning the ε-tree): ongoing research at LRE. A recurrent graph neural network builds the ε-tree, a noise-aware hierarchical structure that estimates a clean image's component tree from its noisy version. PyTorch Geometric, Higra.
- CAGE (Classifier-Aware Generative Enhancement): a U-Net denoiser trained against a frozen classifier. Its outputs reach 0.69 accuracy against 0.57 on the clean images (3 seeds), and Grad-CAM shows the class-discriminative patterns it adds beyond denoising.
- KHUNC (Kinda Half U-Net Compression): image compression with a split U-Net, evaluated with SSIMULACRA2 on the 24 Kodak images.
- SingularityRenderer: physically based rendering of a rotating (Kerr) black hole and its accretion disk, with RK4 geodesic integration on CPU and CUDA. With Valentin Best.
- ZeFauxRest: real-time volumetric forest in C++17 and OpenGL 4.5, with an 8-pass deferred pipeline, ray-marched clouds and godrays. With Martin Boulanger.
- OMaR (Optical M(a)usic Recognition): score transcription with no trained classifier, using directional morphology and graph rewriting to go from printed scores to MIDI. 0.51 mAP on DeepScoresV2, and the C++20 port is 6.1x faster than the Python reference. With Sacha Melin.
- Slopetimization: comparison of gradient descent, conjugate gradient, Newton and BFGS on convex and non-convex functions. With Sacha Melin and Valentin Best.
- Soratra: a self-hosted reading journal for a small circle of friends.