Projects
A few things I've built and studied. Where the code is public it's linked; some of it is proprietary or runs on live credentials, so it's described rather than shown.
Systems & startups
- Kalshi market-making botNov 2025–present
A fully automated market-making engine I run on prediction markets: real-time order-book ingestion, automated quoting, and inventory control. Across two live accounts it booked ~$64K through June 2026 at a 5.9 annualized Sharpe, green on 88% of trading days.
Solo projectRead more →
- AxisOct 2025–present
Co-founder and CTO of a Y Combinator–backed startup in San Francisco. I wrote 95% of the commits: a Python ingestion pipeline, a provenance-tracked knowledge graph over Postgres, and a TypeScript forecasting backend powering an agent-assisted research dashboard for commodity markets, where agents do the legwork and analysts stay in the loop. The product reached ~$60K in annual recurring revenue.
Co-founder & CTORead more →
A mobile social app I built and shipped (Flutter + Firebase) that grew to 500+ users across a few local high schools, with real-time feeds, anonymous posting, polls, and DMs. Launched with an App Store waiting-room trick and an iPad giveaway funded by Super Bowl betting promos.
Primary developerRead more →
Research
I gave two frontier LLMs, GPT-4.1 and Terra, the same sunk-cost and lottery decisions as a chooser, a financial advisor, and a predictor of human behavior. GPT-4.1 split sharply by role; Terra did not. The awkward part: the advisor was not reliably rational.
Independent researchRead more →
- Evaluating language models on human behaviorSpring 2025
My senior thesis. I designed an evaluation suite to test whether language models can predict how people answer survey questions, with normalized accuracy metrics, uniform and distributional baselines, and a seen/unseen split. A dummy-option ablation separated 'learning the format' from 'learning the behavior,' and turned up a negative result.
Senior thesis · advised by Dr. David van DijkRead more →
A study of how gradient-based optimizers hold up when a fraction of the training data is adversarially corrupted. I compared Robust Gradient Descent against SEVER, worked through the theory, and ran simulations across contamination levels, including a hybrid of the two.
Course research · CPSC 561Read more →Source →
A physics-research method for measuring superconducting energy gaps from photoemission data. Instead of fitting one energy slice at the Fermi momentum, a two-stage '1.5D' fit reads the gap off the whole band's dispersion, which cuts fit residuals several-fold (up to ~20×) when the gap is smaller than the instrument's resolution. 2nd-place poster at a Berkeley Advanced Light Source user meeting.
Yale Applied Physics · advised by Prof. Yu HeRead more →Source →
Explorations & writing
Two people take turns crossing restaurants off a list until one is left. A short Monte Carlo essay on why the nightly dinner negotiation is secretly a mechanism-design problem, and why cutting the list by fractions beats cutting it by fixed numbers when nobody has the patience for more than a few rounds.
- How to look taller in a mirrorMarch 14, 2021
Stores lean their mirrors back because it makes you look taller. Finding the best angle sounds like a one-line optimization until you try to pin down what "taller" even means. It turns out to be at least three different things, and they disagree. An optics puzzle that started in a Uniqlo fitting room in 2021.