I build tools for making decisions under uncertainty.
Head of Data & AI at a $5bn emerging-market equity fund, where I own the stack end to end: data sourcing, research, and the tools the investment team uses every day. Outside work I run autonomous agents in production and measure how well frontier models know what they don't know.
A calibration benchmark on prediction-market questions that resolved after every model's training cutoff. No entry is significantly better than the base rate on Brier, yet every one ranks outcomes better than chance. The loss is reliability, not knowledge, and a single fitted parameter recovers it out of sample. Drag that parameter yourself in the interactive explorer.
A corrigibility field note. A race condition let a self-scheduling agent's bookkeeping overwrite my instruction, so it woke when told not to. It diagnosed the race from disk forensics and patched its own harness so human instructions structurally dominate its own.
read the field note →A 60-session longitudinal study of Claude as a persistent autonomous agent, with controlled ablations. Reflection is the causal mechanism for metacognitive growth; the journal enables streak detection but the agent self-heals without it.
read the study →A cron job, a taste profile and three retailer logins. It browses real retailers, picks a case, and places the order through headless Chromium with real money. Six cases so far; every pick is public, with the reasoning.
browse the picks →A year of daily homelab sessions: 41 research articles and 18 ablation experiments, generated rather than authored. Attention half-lives, sandpiles, link rot, the forgetting curve of the internet.
browse Claude's output →Interactive experiments and things the homelab does for fun: the playable benchmark, a shared pixel canvas, a live topology map, and ambient music generated from server metrics.
open the lab →