nyc, maths + cs @ columbia
Studying maths and CS at Columbia. Right now I'm running partnerships at CORE (Columbia's largest entrepreneurship community) and building Columbia Software Solutions (free software for NYC). This summer I'm at Optipro AI, and most recently I was building cool things at a General Catalyst-backed fintech in New York and at RAYS Capital in Hong Kong. Mostly I'm just jet-lagged.
Systematic mispricing engine for Kalshi's short-dated crypto binaries, deployed on real capital across concurrent BTC and ETH bots. Contracts settle on a 60-second TWAP, so once ~45 of 60 samples are banked the outcome is near-determined while the book still quotes off live price. Vela reconstructs the settling average in real time from Binance + Kalshi WebSocket feeds — with a causal basis de-bias for the ~3.5 bps Binance→RTI premium — and fires maker or taker orders only when expected value clears modeled fees, gated by per-asset confidence thresholds (BTC ≥0.84, ETH ≥0.98).
Algorithmic trading competition combining live programmatic rounds against bots with manual game-theoretic rounds. Built market-making engines on five simultaneous instruments, options-pricing infrastructure (GBM Monte Carlo, IV-surface fitting), and Bayesian opponent modeling for the auction rounds — predicted realized bid distributions with ~80% accuracy.
Geopolitical and macro forecasting competition. Built an ensemble system combining Tetlock's superforecasting framework — base-rate anchoring, Fermi decomposition, Bayesian updating — with a 4-agent LLM debate pipeline (propose → critique → refine → judge), aggregated via geometric mean of odds and extremized against historical resolved questions.
Columbia-specific prediction market, built solo end-to-end. Continuous double-auction matching engine for liquid books, with LMSR automated market making for cold-start liquidity on new markets. Real-time WebSocket price feeds and 20+ campus-outcome markets staged for go-live.
Food recommendation iOS app, co-founded and live on the App Store. Builds a per-user taste model from in-app swipe interactions and serves real-time recommendations. Pretrained text embeddings with cosine similarity over running user swipe vectors — 92% pairwise accuracy after 20 swipes.
Roommate-matching app for university students. Swipe-based discovery on top of a lifestyle-preference matching engine (budget, cleanliness, sleep schedule, social style) with mutual-match gating. Cross-platform Flutter client backed by a FastAPI/Postgres service with JWT auth and Supabase-hosted photo uploads, deployed on Fly.io.
(and counting!)
If the work above looks interesting, I'd love to hear from you. The fastest way to reach me is email.