ALEX
WU.
I build systems that know when to refuse.
Any system can be made to answer. The rarer one knows its own edges: it abstains when unsure, and says no when no is the honest reply. Calibration over confidence. That is the discipline.
Crucible1st Place · Datadog Hackathon
Adversarial evaluation for agentic systems. Measures where models break under pressure, not where they behave.
Lacuna3rd Place · Mistral Worldwide Hackathon NYC
A cross-lingual divergence explorer. Maps where a model understands the same concept differently depending on the language you ask in.
PsychohistoryML
Survival analysis of historical institutions across 256 polities. The headline result did not survive a leave-one-era-out holdout, so it was retracted rather than published.
OpenZerg
Evolutionary adversarial red-teaming. An attack registry with a fitness controller that learns to break the target.
Sixth BoroughWinner · NVIDIA Spark Hack
Real-time 3D of ~45,000 NYC buildings from open footprint data. The skyline you are flying through here is its descendant.
Morel CompassHonorable Mention · Flight Zero
On-device AI that refuses rather than confidently confirm a poisonous mushroom. The refusal is the product.
Majikari
A Japanese resale-market deal detector. Ingests and classifies around 700,000 listings, runs unattended on a daily schedule, and watches itself well enough to tell a real zero from a blocked read.