I build models that work.correction: tests that tell me when my models don't

Trying to catch cheating in coding assessments without accusing anyone innocent, which turns out to be a statistics problem more than an engineering one. The most useful thing I built was the test that showed a promising 0.778 AUROC was indistinguishable from chance, and a plain lookup of which question a student got beat it anyway.

I built and ran the platform for my residence's Emerald Challenge alone, start to closing ceremony. Ninety-six teams played, and the review queue was empty when it closed.

Six months building AI tooling for a team where the usual shortcuts, like pointing Copilot at a problem, aren't allowed. I'm prouder of the guardrails I built to check every action an agent takes than of the agents themselves, and of a Snowflake-and-SharePoint pipeline that turned hunting for audit evidence into something close to instant.

A team final-year project, advised by Anand Bhojan, that turns a plain-English trading idea into a backtest for students who haven't yet learned a strategy can cheat by knowing the future.

Hand and eye tracking on Apple Vision Pro and iPhone for stroke rehabilitation, in collaboration with Apple. I built the visionOS side and the iPhone companion app, and it was tested with real clinicians and real patients. A paper is in preparation and a patent application is in progress.

A research engine watching Polymarket and Kalshi for the same question priced differently on each. It never places a trade, since prediction markets are restricted where I live.

Teaching CS3230, the algorithms course I took myself as a student, and CS3263, foundations of AI. Four semesters in, students rate the teaching 4.8 out of 5, and it got me onto the NUS Honour List of Student Tutors.

I wanted to know how small a BERT question-answering model could get before it noticeably got worse, and where exactly that line was.