Projects

Can LLM agents, one per household, negotiate in plain English to fairly share limited solar and battery power during a grid outage? In a deterministic 30-household simulation, live agents beat a zero-LLM control by 5.8 points of served load on every clean seed, closing 29% of the gap to a perfect-information oracle, with fairness improving alongside. With Prof. Yongfeng Zhang, Rutgers CS.

LSTM neural networks predicting stock prices from Twitter sentiment: 80,793 tweets across AAPL, TSLA, and MSFT with 13 technical indicators. Adding sentiment raised average RMSE by about 32%, a result that runs against common assumptions in financial ML. Accepted by the Journal of Emerging Investigators.

Web app that turns handwritten notes into study resources: OCR (Google Cloud Vision + PyMuPDF), Claude summarization, a social layer with follows and comments, and a student marketplace. 100+ routes, 30 models, OWASP-aligned audit logging.

Top-down wave-based arcade shooter built around a time-rewind mechanic: record 1.5 seconds of movement, then rewind at double speed to damage enemies along your trail. Mini-bosses, chain-kill multipliers, and a top-5 high-score table.