Things I've built — software, ML, and the occasional experiment.
Reinforcement-learning-based red teaming for LLM applications: adaptive multi-turn probing, prompt mutation, and MITRE ATLAS-mapped findings, with JSON, HTML, and SARIF reports for security review.
Quant pipeline extracting timestamped trading signals from SEC filings (10-K, 10-Q, 8-K) via LLMs — grounded in Cohen, Malloy & Nguyen (2020). Four engines: supply chain graph, risk factor diff, M&A detector, and tone tracker. Outputs to Parquet/DuckDB with FastAPI and webhook support.
Differentiable physics library covering classical mechanics, FDTD electromagnetics, quantum mechanics, optics, and statistical mechanics — fully differentiable end-to-end. Enables inverse problem solving via autodiff.
A predictive F1 race winner app using 70+ years of historical data and advanced machine learning models for 85% prediction accuracy. Processes 1,000+ data points per race, analyzing factors like driver form, car performance, and track familiarity. Deploys on AWS/GCP with 99.9% uptime.
A real-time ASL translator built with Convolutional Neural Networks (CNNs) for gesture recognition and TensorFlow for model training. Processes webcam inputs to detect and translate ASL alphabet signs (A-Z) into text with high accuracy.
An advanced web application utilizing CNNs for facial recognition and spectrogram analysis combined with the Levenshtein distance algorithm for speech pattern detection. Designed to assess sobriety in real-time to prevent drunk driving through ML-powered evaluation.
