ML/AI Engineer · École 42 Paris · 42AI Lab Builder

I build ML systems from the ground up — implementing architectures from scratch rather than wrapping libraries, with a focus on understanding what happens at the tensor level. Currently deepening my work on LLM internals and neural signal processing.

📖 Studying LLM internals — attention, inference, architecture tradeoffs 🔧 Building and publishing ML/DL projects

I'm fascinated by the gap between using AI and understanding it. Most people can call an API. I want to know what happens inside — how attention routes information, why a tensor is or isn't contiguous in memory, what a model is actually doing when it "reasons." That curiosity is what pulls me from one project to the next, and it's why I learn fastest when I'm building at the lowest level I can reach. I'm convinced that understanding what happens inside AI is the key to better use and better models.

My goal: to work on the systems that push what these models can do — and to be the kind of engineer who understands them from the math up, not just the API down.

📊 weight-your-wage — A full end-to-end MLOps pipeline predicting developer salaries from the Stack Overflow survey. Data ingestion to MinIO, ETL into PostgreSQL, training with PyTorch Lightning tracked in MLflow, real-time inference via FastAPI, dashboards in Metabase, and full observability with Grafana/Prometheus/Loki — all containerized with Docker Compose. The project where "deploying a model" stopped meaning a single script.

🔢 Multilayer-Perceptron — A neural network built entirely from scratch in NumPy — no framework. Forward pass, manual backprop, and my own autograd engine that computes gradients automatically, the way PyTorch does. Modular by design: layers, activations, and optimizers as composable pieces. The project where deep learning stopped being magic.

📊 DSLR — Multiclass logistic regression with no scikit-learn for the model. Batch, stochastic, and mini-batch gradient descent, implemented and compared by hand.

🌐 ft_transcendence — Real-time multiplayer web app with WebSockets, OAuth2, and a Dockerized microservices backend.

🧠 [total-perspective-vortex] (publishing soon) — A brain-computer interface pipeline. Raw EEG → signal processing with MNE-Python → classification. Understanding intelligence by reading it off the cortex.

🤖 GPT-2 from scratch (publishing soon) — A transformer rebuilt in PyTorch: multi-head attention, residual streams, the full (B, T, C) shape flow. Built to understand, not to ship.