About
I'm currently completing my Master's degree at Télécom SudParis (Institut Polytechnique de Paris), specializing in data science and applied mathematics. My work spans computer vision, NLP, probabilistic modeling, and mathematical optimization.
I don't just run .fit().
I build the thing, understand every gradient, and then trust it.
From implementing backpropagation by hand in NumPy to fine-tuning BERT on real data,
I care about what happens between the math and the result.
When I'm not training models, you'll find me trail running in the Alps, playing jazz piano, or planning the next mountaineering trip.
Projects
Mushroom Toxicity Classifier ↗
End-to-end ML pipeline on categorical data: MCA encoding, hierarchical clustering, LDA classification, model benchmarking. Interactive Streamlit dashboard, CI/CD via GitHub Actions, Makefile-orchestrated pipeline.
CNN Explainability Workshop ↗
What do CNNs actually learn? Filter visualization on VGG16, Grad-CAM heatmaps, occlusion sensitivity, a quantitative challenge of CNIL anonymization guidelines, and one-shot face recognition via transfer learning.
Deep Learning from Scratch ↗
From a single neuron to convolutional networks on MNIST — 12 notebooks, 3 PDF guides, every gradient derived by hand. Built as a learning-in-public series with accompanying LinkedIn posts.
NLP: From Scratch to BERT ↗
The full NLP progression in 4 notebooks: text preprocessing, Bag-of-Words & TF-IDF for spam classification, Word2Vec/FastText embeddings, and BERT fine-tuning with Hugging Face Transformers.
Differentiable Particle Filter ↗
Research project implementing differentiable particle filtering via entropy-regularized optimal transport. Full mathematical proofs and auto-differentiation for parameter estimation.
Autoencoders & VAE ↗
Denoising convolutional autoencoders and variational autoencoders on MNIST. Exploring latent spaces, the reparameterization trick, and generative modeling with modular PyTorch code.
Clinical Gait Analysis ↗
Unsupervised analysis of clinical gait data: Dynamic Time Warping, K-Medoids clustering, and 3D feature engineering to study speed constraint effects on knee flexion patterns.
Skills
Education