case studies — problem, constraints, decisions, outcome.
A generative-AI pipeline for finding cryptic pockets in proteins.
#comp-bio#generative-ai#agritech
Rather than watching microsecond molecular simulations, STROBE flashes a protein’s thermodynamic landscape into discrete 3D snapshots and catches hidden binding sites as they open.
Status: passed its first calibration gate on the standard benchmark (27/43 judge-verified captures). Now hardening before the agritech pilot — plant metabolism, fungal, and crop-pest targets.
Gray-box negotiation experiments measuring what a model says vs. what it computes.
#ai-safety#evals
Measuring the gap between what the model privately computes and what it publicly says. (Spoiler: it knows exactly what it’s doing.)
Status: write-up in progress.
Custom JIT compilation and fused-kernel development for small-tile matrix workloads.
#jit#kernels#apple-silicon
Beating Apple’s Accelerate framework where the tiles are small — ~5x at tile-sized GEMM versus the baseline, with a synthetic-data pipeline for benchmarking against realistic agent traces.
Status: live thread. The compute core under Dawn State.
Research-grade machinery on top of physical corn markets.
#corn#markets#quant
Twenty years of term structure, seasonal cones, carry regimes — better calls at the farm gate. And agent-based modeling of the markets I trade: farmers, elevators, and traders as autonomous agents, letting macro behavior emerge instead of assuming it.
Homework: Empire of Cotton: A Global History — Sven Beckert. Four centuries of the original commodity supply chain.