Philadelphia · Singapore · Jülich
i like to build things.
usually somewhere between machine learning, markets, and the brain.
I'm a sophomore at Penn studying Computer Science at SEAS and Economics & Finance at Wharton. I like problems that sit between fields, and once something grabs me I have a hard time putting it down until it actually works.
Lately that's been AI agents, EEG models of the brain, and a bit of quant on the side. I'm pretty relentless once I'm into something, and honestly I just really like building things that work.
Built an ad-optimization pipeline (Python, Meta and Google Ads APIs) that generates LLM media and A/B tests it across products. Added multi-armed bandit budget allocation that drops underperformers on its own, for a projected 50% lift in ROAS, inside a loop that feeds results back into the next round of creative.
Building an agentic AI system with a 45-tool registry that takes real CRM actions like creating deals, moving pipeline stages, and drafting follow-ups, all behind confirmation gates, snapshot undo, and trust-gating for anything destructive. Also built the RAG layer on Postgres/pgvector and a routing layer across 7 LLM providers with fallbacks, consensus voting, and per-call cost tracking.
Started a freelance web dev agency serving small businesses across Asia, hired and trained a team of five, and shipped 20+ production sites and apps with auth, Stripe, and Firebase/Supabase in React, React Native, and Vue.
Built one EEG pipeline (Python, MNE) across 8 datasets, 504 people, and 8 clinical conditions. Fit a single microstate model, pulled 24 features per person, and showed the features separate neurodegenerative disease from controls, with an Alzheimer's model that transfers to unseen FTD. Ran the paranoid checks too: age regression, 5,000-shuffle permutation tests, cross-site validation, and ComBat, which caught a spectral confound faking a depression result.
Benchmarked a range of LLMs on multimodal sarcasm detection under Prof. Cheston Tan, then fine-tuned a custom model on what the benchmarks showed to sharpen context and text-image understanding.
A measurement-first research platform for prediction-market microstructure. Headline finding: at scale the market is efficient — proving it took instruments that caught nine measurement artifacts standard backtesting would ship as edges, each documented with receipts across an 880k-market dataset. A cross-venue benchmark against equity TAQ shows wide prediction-market books booking a "realized spread" ~80× a small-cap maker's take. Zero-server self-healing CI at $0 infra, tests green. Two working papers.
A from-scratch LLM inference server, no custom CUDA, and the interesting part is the scheduler. I built the batching ladder from a naive loop up through continuous batching and a paged KV cache, for 9.6× the throughput at a fraction of the latency, reaching 16% of vLLM on the same GPU. Then I turned it on published optimizations as an audit: every engine is checked token-for-token against a baseline and measured past a noise floor, which killed five of my own first-pass results.
A quant research engine built to attack its own signal before trusting it. Vol-targeted backtests, walk-forward validation, and 70+ tests in CI. I rebuilt a point-in-time S&P universe to measure survivorship bias directly, found most of the edge was a backtest artifact, then recovered a smaller real one at half the cost.
One pipeline across 8 public datasets and 504 people. It fits a single microstate model, pulls 24 features per person, and classifies brain disorders from a few minutes of resting-state EEG. Alzheimer's separates from controls at 0.89 AUC, and a model trained on it transfers to FTD it never saw.
A kids' app with its own economy. Kids earn minutes for chores and routines, parents approve requests and stock a reward store, and a dashboard tracks progress. Built in React Native with Firebase, and live on the App Store and Google Play.
Python · TypeScript · JavaScript · C/C++ · Java · SQL
PyTorch · scikit-learn · pandas · NumPy · XGBoost · Hugging Face · MNE-Python · RAG (pgvector)
React · React Native · Node.js · Express · Next.js · Prisma · REST
PostgreSQL · Redis · BullMQ · pg-boss · Supabase · Firebase · Docker
Data Structures & Algorithms · Computer Systems · Automata & Computability · Probability · Linear Algebra · Discrete Math