Thibault Castells
I turn large generative AI models into faster, smaller versions that can run under real device constraints.
Core areas
Selected work
Work across on-device generation, diffusion and LLM pruning, model-compression tooling, deployment workflows, and curriculum learning: EdgeFusion, LD-Pruner, Shortened LLaMA, BK-SDM, NetsPresso, automatic pruning, and SuperLoss.
On-device text-to-image generation system targeting mobile NPUs with optimised diffusion models.
Achieved sub-second inference budgets through graph-level optimisations and quantisation.
Experience
Research engineer at Nota AI, specialising in neural network optimization.
Technical writing
Code-first articles on model training, inference, optimization, and practical AI engineering. Tutorials will appear here as they are published.
Profiling and optimizing a small autoregressive transformer with JAX, KV caching, batching, graph compilation, and low-bit inference.
A code-first experiment that builds a character-level decoder-only transformer with JAX, Flax, and Optax, then trains it on generated addition, subtraction, multiplication, integer division, and modulo problems.
Side projects
Plume is a local Markdown app for notes, lists, and project planning. I build it as an independent product outside work.
