CoDHe LabsInsights, code, and write-ups from Daniel Halwell’s CoDHe Labs — rigour-first open work where AI meets chemistry.https://codhelabs.co.uk/From SMILES to graph transformers — designing a teaching curriculumhttps://codhelabs.co.uk/notes/2026-05-17-chemical-graph-curriculumhttps://codhelabs.co.uk/notes/2026-05-17-chemical-graph-curriculumA seven-notebook open-source curriculum that takes a chemist or ML engineer from molecular graph basics through positional encodings, attention, graph transformers, equivariance, and property prediction on real benchmarks — with notes on why the lessons are ordered the way they are.Sun, 17 May 2026 00:00:00 GMTWhen the experiment matters as much as the molecule: multi-modal property predictionhttps://codhelabs.co.uk/notes/2026-05-17-multi-modal-qspr-when-the-experiment-mattershttps://codhelabs.co.uk/notes/2026-05-17-multi-modal-qspr-when-the-experiment-mattersSome chemical properties — HPLC retention time being the obvious one — are functions of the molecule and the experimental conditions together. A multi-modal framework that fuses molecular graphs, tabular descriptors, and experimental context, with notes on architecture choices and fusion strategies.Sun, 17 May 2026 00:00:00 GMTDesigning a small Mixture-of-Experts that actually routeshttps://codhelabs.co.uk/notes/2026-05-17-small-moe-that-actually-routeshttps://codhelabs.co.uk/notes/2026-05-17-small-moe-that-actually-routesA ~363M-parameter, weight-shared MoE language model designed so the router cannot quietly collapse to a uniform distribution — architecture, routing algorithm, balance loss, optimiser ablation, and what to watch in training.Sun, 17 May 2026 00:00:00 GMTFrom mass spectra to molecular graphs with spectral diffusionhttps://codhelabs.co.uk/notes/2026-05-17-spectral-diffusion-mass-spec-to-graphhttps://codhelabs.co.uk/notes/2026-05-17-spectral-diffusion-mass-spec-to-graphA spectral diffusion model that conditions on tandem mass spectra and generates the Laplacian spectral embedding of the underlying molecular graph — design choices, identifiability traps, and the projection-aware trick that makes the targets stable.Sun, 17 May 2026 00:00:00 GMTUncertainty as a compute budget: a small adaptive language modelhttps://codhelabs.co.uk/notes/2026-05-17-uncertainty-as-a-compute-budgethttps://codhelabs.co.uk/notes/2026-05-17-uncertainty-as-a-compute-budgetSpending more compute per token only when the model is unsure — a recursive sparse transformer with sliding-window attention, multi-signal uncertainty estimation, and uncertainty-triggered reasoning.Sun, 17 May 2026 00:00:00 GMT