I’m a senior at UCLA studying Statistics & Data Science with a Data Science Engineering minor, graduating Spring 2026. I work at the intersection of machine learning, causal inference, and quantitative trading — the common thread is building systems that have to survive contact with real, messy, adversarial data.
My current focus is neural game-solving and experimentation methodology. On the poker side that means scaling Deep CFR from Leduc through HUNL into 5-card PLO — composition-dependent encoders, batched sigma schedulers, and honest exploitability evaluation. On the experimentation side it means switchback designs, deflated Sharpe ratios, walk-forward backtesting on marketplaces and prediction markets, and quasi-experimental work — interrupted time series, difference-in-differences, hierarchical Bayesian models — applied wherever the data is dense enough to identify something real.
I also build trading systems. The throughline of all of it is anti-overfitting: frozen out-of-sample holdouts, parameter-neighborhood robustness checks, conservative cost models, and minimum-trade floors — because the easiest thing in the world is to fabricate a backtest.
Outside of code I play tournament pickleball, snowboard, and grind heads-up cash. Each of them rewards the same thing my work does: cold expected-value reasoning under noise.
Selected work
- Subscription churn + causal uplift — LightGBM AUC 0.866 on real KKBox data (6.8M members); R-learner retention targeting nets +92% more value than risk-targeting on the same budget.
- Switchback experiment simulator — diagnose marketplace interference; recover a 200% biased estimate.
- Pro Valorant prediction models — 28K matches, 600K player-maps; GBM beats Elo on match outcomes, series-kills prop beats naive by +9.6% (>36σ), with the leakage story that killed an earlier “17% lift”.
- Hierarchical Bayesian Valorant ratings — Bradley–Terry with map / region / time pooling; beats Elo by 0.040 log-loss on held-out cross-regional matches and cuts ECE 0.092 → 0.060.
- Valorant Patch 5.12 causal study — ITS + DiD across maps; −78pp Chamber pick rate, +0.42 bits sentinel entropy, four robustness checks aligned.
- Deep CFR for HUNL & PLO5 — full neural game-solving pipeline, 200-iter HUNL training, K=10k PLO5 training.
- Polymarket research toolkit — scraping, walk-forward backtester, deflated Sharpe, complementary-pair arb.
- Crypto strategy discovery — BTC/ETH walk-forward research with frozen OOS holdout.
- LEAP trading strategies — long-dated leveraged option backtests, drawdown stability, self-funded variants.
- Pickleball CV — YOLOv8 + ResNet50 keypoint pipeline with homography-driven analytics.
- ORB algorithmic day-trading — XGBoost-gated TQQQ breakout strategy, +19.1% annualized lift.
- No-Bust 21st Century Blackjack Monte Carlo + CDZ⁻ solver — composition-dependent strategy for a California card-room variant.
See the full set on the Projects page.