Projects

Code, tools and demos: AI for mathematics, generative models, ML for science, and LLM systems and agents.

AI for mathematics

Human–AI Mathematics
An open framework for mathematical research with AI agents, in which a claim counts as proved only after an independent review. Its first case study is a manuscript that reconstructs and compares the three proofs of the KLS conjecture. Post

latexfmt
A LaTeX formatter that understands mathematics. uv tool install latexfmt · PyPI

Generative models

edm-error-propagation-code
The experiments and archived runs behind the EDM error propagation preprint: local errors of the first-order EDM sampler and the amplification they actually undergo. arXiv:2610.10190

flow-matching-notes
The notebooks behind the flow matching posts: flow matching as posterior averaging and closed-form flow matching.

ML for science

mhcpred
Predicting MHC class I peptide presentation with a neural network, benchmarked against MHCflurry. Post

raman-spectra
Preprocessing Raman spectra and predicting analyte concentrations with PLS regression. Post

LLM systems & agents

haiku-shunt
A Claude Code plugin that delegates large file reads to a small model, and an A/B test of what that actually saves.

llm-slide-deck
Rebuilding a real slide deck with an LLM in four formats, and where each one fails. Post

llm-observability
End-to-end tracing of a FastHTML chatbot with Langfuse. Post

pdf-rag
Retrieval over a corpus of PDFs: recovering document structure with a vision language model, extracting metadata, and querying with a two-level ReAct agent. Post · demo

pdf-parsing
A comparison of Docling, Marker, MinerU, pymupdf4llm, LlamaParse, Gemini and Mistral OCR on real documents. Post · demo