Brad Cao

Awarded top poster recognition at one of the nation's premier clinical pharmacology conferences, for research on predictive modeling of serious adverse drug outcomes in stimulant-treated patients.

Placed 1st nationally in the Data Analysis event at the Future Business Leaders of America National Leadership Conference, competing against thousands of students across the country.

Received a formal citation from the U.S. Congress in recognition of outstanding community contributions and achievements.

Named to the USA Fencing 1st Team All-American squad in Men's Epee, one of the highest individual honors in U.S. competitive fencing.

Achieved Silver division standing in the USA Computing Olympiad, a highly selective national algorithmic programming competition.

Conducted clinical pharmacology research under faculty mentorship, working with over a decade of FDA adverse event reporting data. The project focused on identifying ADHD patients at elevated risk of serious adverse outcomes from stimulant treatments using real-world evidence. Led the full data pipeline design, feature engineering from raw FDA submissions, and the development and optimization of a random forest classification model. Findings were presented as an award-winning poster at the 2025 ASCPT National Conference.

Cao, B., Lee S., Naga A., Gobburu J., Dunn A.

Cao, B., Gupta P., Dunn A.

An event-driven limit order book simulator with exact price-time priority, built to test how much of a backtested edge survives realistic queue position. Most backtests fill your order the moment a trade prints at your price; real exchanges make you wait.

A strategy worth +214 ticks per episode under the standard fill assumption is worth +2.8 once the queue is modelled — 98.7% of the edge was the assumption rather than the strategy. The synthetic market is validated against published stylized facts (9 of 11 inside empirical bands), and every reported result carries paired bootstrap confidence intervals, Holm–Bonferroni correction, and a deflated Sharpe charged for all 24 configurations tried during development.

Reconstructs per-wallet profit and loss from Polymarket on-chain activity and validates it against the exchange's own numbers: 92% of positions match exactly and 43 of 51 wallets reconcile perfectly.

It then asks whether apparent trader skill is real or variance. Population raw edge of +0.126 collapses to −0.095 once shrunk toward the mean, so most of the visible edge is luck. Backtested point-in-time, the skill-weighted signal does not beat the market price — Brier 0.2088 against the market's 0.2046, ROI −3.3% after costs. Reported as-is.

An empirical study of double descent: the observation that past a certain size, bigger models stop overfitting and start improving again. Sweeping a random-feature model through the interpolation threshold on MNIST, test error peaks at exactly p = n — 88.8% there against 24.4% at five times the size.

The mechanism is measured four independent ways from a single matrix decomposition per model: the fitted weights grow 313×, the feature matrix's smallest singular value collapses, a bias–variance decomposition puts 90.0% of the peak on variance, and a ridge penalty tuned per model size removes the peak entirely. Ensembling, which should reduce variance, fails at the threshold — the members' output scales differ by 10.5× there — and the repository reports that with its cause rather than omitting it.

A mechanistic-interpretability study of how language models learn from their own context. An induction head is an attention head that looks back for the last time it saw the current token and predicts whatever followed it. This scores all 144 attention heads in GPT-2 Small, locates them, and switches them off.

With the top four ablated, loss on a repeated random sequence rises from 0.23 to 7.83 nats, while ablating four matched control heads from the same layers leaves it at 0.27 — a separation of 86 standard errors. On aggregate in-context learning over natural prose the same intervention is much weaker and not statistically decisive, which the write-up states plainly rather than leading with the more flattering measurement.

MyCureHope is an open-source disease treatment database built to make medical information more accessible to patients and families worldwide. Rather than requiring users to know their exact diagnosis, the platform allows people to search for potential treatments based on symptoms — a more intuitive approach to navigating complex healthcare information.

The platform now serves 33,000 users across 78 countries. It integrates data aggregation algorithms that compile and organize treatment data from multiple trusted sources, helping individuals identify relevant options and better understand the pathways of care available to them.

Competing nationally in Men's Epee on the North American Cup (NAC) circuit. Ranked Top 30 among cadets nationally (2024 season) and Top 50 among juniors (2025 season).