Engineering student · Python developer @ Framatome · France
Mathis
Giovannetti
AI is the biggest revolution since we learned to make fire. I intend to help build it.
§1 — Who
Engineering student at Centrale, apprentice Python developer at Framatome, and member of Automatants, the school's AI association. I learn by rebuilding things myself. So far that means a tokenizer, attention layers, a full translation Transformer and GPT-2, written line by line until none of it feels like magic.
§2 — Now
- Python development apprenticeship at Framatome
- Re-implementing modern transformer variants (RoPE, RMSNorm, SwiGLU)
- Building LLM agents with LangChain and local models
§3 — Next
Work in AI: contribute to, and take part in, the biggest revolution in human history since the discovery of fire.
Selected work
English → French Transformer, from scratch
A full encoder–decoder Transformer in PyTorch for EN→FR translation, built twice: once like the 2017 paper, once with the modern recipe (RMSNorm, SwiGLU, RoPE).
Nov 2026 ★★★ N°003GPT-2 from scratch
Re-implemented GPT-2, loaded the official 124M weights into it, and fine-tuned it into an SMS spam classifier. Also coded a tokenizer by hand.
Nov 2026 ★★★ N°001LangChain agents with local LLMs
Tool-using agents, streaming, runtime context and dynamic prompts with LangChain, all running locally on Llama 3.2 through Ollama.
Sep 2026 ★★★The index
★ Exploration★★ Build★★★ Deep dive
- N°004 Nov 2026 English → French Transformer, from scratch in progress A full encoder–decoder Transformer in PyTorch for EN→FR translation, built twice: once like the 2017 paper, once with the modern recipe (RMSNorm, SwiGLU, RoPE). ★★★ →
- N°003 Nov 2026 GPT-2 from scratch Re-implemented GPT-2, loaded the official 124M weights into it, and fine-tuned it into an SMS spam classifier. Also coded a tokenizer by hand. ★★★ →
- N°002 Oct 2026 CIFAR-10 ResNet challenge Automatants weekly challenge: from a baseline CNN to a ResNet-18 with skip connections, tuned to climb the CIFAR-10 leaderboard. ★★★ →
- N°001 Sep 2026 LangChain agents with local LLMs Tool-using agents, streaming, runtime context and dynamic prompts with LangChain, all running locally on Llama 3.2 through Ollama. ★★★ →
Nothing matches. Loosen the filters.