Skip to content

~/thibaultlab

Lab

Experiments, side projects, and things I built to understand something better. Not everything here is finished — that's the point.

  • GPT-style language model from scratch

    building

    A decoder-only transformer in PyTorch, to understand what the libraries hide.

    llm · transformers · pytorch

    fig. 1 — causal attention

  • Self-hosted LLM inference (home lab)

    building

    Open-weight models running on my own hardware, sized to fit a 12GB GPU.

    llm · self-hosting · quantization

    fig. 2 — one 12 GB budget

  • CV Generator: an LLM CV reformatter with an evaluation harness

    done

    Turns an existing CV into a clean LaTeX PDF, and measures where the LLM extraction breaks.

    llm · evaluation · nlprepo →

    fig. 3 — extract · render · check

  • Solving sudoku with reasoning, evolution and deep learning

    done

    One puzzle family attacked with SAT solving, an evolutionary algorithm and a transformer, then compared.

    sat · evolutionary-algorithms · transformersrepo →

    fig. 4 — one grid, three solvers

  • BSc thesis: classifying helpdesk tickets with little labeled data

    done

    Supervised vs. unsupervised approaches to classifying unstructured helpdesk tickets under low-resource conditions.

    nlp · classification · clusteringrepo →

    fig. 5 — labeled vs. clustered