AI Infrastructure & Systems — ENSTA Paris, 2nd year
ChaosAI — Time-series world model trained from scratch.
- 38M-parameter Mamba-2 JEPA encoder, trained on 838M tokens across 8,969 financial assets
- Full JAX/Flax pipeline: FSQ tokenizer → SSM encoder → OT-CFM stochastic predictor → TD-MPC2 RL agent
- Auto-sharding on TPU v6e clusters (GSPMD, 2D mesh, XLA production flags)
- Data lake: raw parquet → ArrayRecord on GCS, zero idle cost
The core insight: JEPA (Joint Embedding Predictive Architecture) learns structured latent representations of relationships and context — not next-token prediction. Same philosophy as knowledge graphs for agents.
AI/Compute: JAX, Flax, Optax, PyTorch, XLA, TPU Pod topology
Infra: GCP, GCS, Grain/ArrayRecord, Docker, FinOps
Systems: Python, C, Bash