Engineering notes.
Architecture decisions, benchmarks, failure tests, open-source work, and what changed when implementation disagreed with the plan.
Better Protein Representations Were Not Enough: A Fly-Inspired Search That Lost
I tested whether a mushroom-body-inspired sparse code could find high-fitness protein variants with a handful of measurements. With matched budgets, it lost to random search, and the reason was the search, not the representation.
Read →9 min · Single-Cell / Perturbation Models / Evaluation / ReproducibilityCellForge: My Trust Layer Said Everything Was Unreliable, and It Was Wrong
A GEARS run on Norman Perturb-seq data lost to simple baselines, and my own reliability rule marked every held-out perturbation unreliable. The second result was a bug in my method, not a finding about biology.
Read →8 min · gpu / metal / cpp / bioinformatics / performanceMetalSW: When My GPU Code Lost to the CPU
I built exact Smith–Waterman protein search on Apple Silicon. The first serious benchmark lost to Parasail. This is what I measured, what I got wrong, and what eventually pushed the GPU to 13.5× the CPU SIMD baseline at 50,000 sequences.
Read →