Projects
The NMR Facility: an NMR spectroscopy game tutored by Claude Fable 5.1
A browser game built around the real workflow of running an NMR experiment: Dr. Novák, the facility manager, hands the player an unknown liquid, and they carry it through sample prep, locking, shimming, tuning, and acquisition, then phasing, referencing, and integrating a genuine 1H/13C spectrum to identify the compound. Four in-game tutors — non-scripted AI characters running on Claude Fable 5.1, each with its own specialty from basic peak interpretation to solid-state NMR (magic-angle spinning, quadrupolar nuclei, zeolite framework studies) — answer open questions and write fresh quiz material grounded in current literature rather than a fixed script.
Play it →Free Weights, Expensive Everything Else
An essay on what running an open-weight model actually costs, across three buyer profiles — individual, small company, enterprise — set against Vercel's record 62% open-weight token share and the leaked economics behind closed-model pricing. The argument: with weights commoditized, what's left to compete over is who can afford the compute, who operates it safely, and who keeps the data.
Read it on Substack →ScienceLM: a small chemistry model, and the discipline of not believing my own results
A ~427M-parameter chemistry language model, trained end to end on a single consumer GPU. The architecture result: replacing most attention layers with a physics-motivated linear recurrence (a Matrix Product State, with a few real attention layers kept for long-range exceptions) beats a plain transformer 4.17× on perplexity at matched size. The more interesting result is a negative one — a fine-tuning pass that seemed to nearly double factual recall (40.3% → 79.2%) turned out to be memorization of the probe itself, once tested against fresh, untouched facts (30.0%, at or below baseline). That gap is the actual subject of the project: a from-scratch retrieval system built to replace fine-tuning, deliberately evaluated on facts the training data never touched, is in progress.
PDF to Markdown for RAG — pdf-inspector vs Docling
An app that converts PDFs with two open-source engines side by side — Firecrawl's new Rust-based pdf-inspector and IBM's Docling — to find out which one to trust with scientific papers. Spoiler: speed is not the whole story; figures are.
Read the write-up →The AI Model Database — every release since 2022
An interactive dataset of every major AI model release since 2022, presented as a game, a timeline chart, and a searchable database. I maintain it because choosing the right model for an R&D team starts with knowing what the field can actually do this quarter.
Explore it →