I build the system. Then I try to prove it wrong.
Most of my work is private. The audits are what I show.
A model that cannot be falsified is a decoration.
Nothing below is aspirational. If it is not in a file, it is not here.
Languages
ML & Data
Backend
Frontend
Infra & Tooling
NEXUS — ML-first crypto trading system. 17k lines, 12 bugs found, 123 tests. Private.
Signal audit — one feature ate 99.63% of the gradient. Fixed it; AUC moved to 0.5149. Still noise. Reported as two findings.
Rejecting an edge — a hardcoded min(0.98, ask) cap manufactured a +0.76c/share result. Recalculated honestly, it wasn't there.
Attendance automation — Playwright over SSO/MFA for ~97 accounts. Key-sealed credentials, preflight exit before any network call.
ThermoApp (public) — SwiftUI over pycalphad for phase diagrams.
hyperspeed-engine (public) — Java/NeoForge. Plus Android GKI, SELinux and kernel tooling.
I will report the negative result. Off-by-one bugs and wrong constants get the same treatment.