I build AI agent infrastructure at Vertex, and I'm finishing my MS in Computer Science at Columbia. The problems I like are the ones where the answer has to actually run: on real data, on real hardware, with real users waiting.
Most of my work starts as a question I couldn't stop poking at. Three that turned into something:
TAOL scores every LLM-generated patch on four risk signals and routes it to auto-apply, human review, or defer. The ambiguity classifier hits 99.2% accuracy across 500 SWE-bench tasks. Turns out you can predict trustworthiness. You just have to measure the right thing, and an ablation study is what tells you which thing that is.
UrbanistAI blurs faces and text before an image ever reaches storage, using three face detectors in a union ensemble because a false positive costs an over-blurred photo and a false negative costs someone's face. There is no code path that stores an unprocessed upload.
Smart glasses for the visually impaired, 3D-printed parts, about $135 in materials. Sensor fusion for distance estimation and context-adaptive keyframe selection so the thing isn't burning compute on redundant frames. Two filed patents came out of it.
Instrumented every tool-discovery call at Vertex before touching a prompt. That's where the 36.11% token cut came from.
78 tests at Vertex, 144 on TAOL. A suite that catches regressions is what lets you keep changing things.
The Pi and Arduino on my desk report live into this site. When they go down, you watch them go down.
Three papers and two patents, every one from an artifact I actually built and ran.
// off-hours: iced mochas, running, tennis, and whatever's half-built on my desk