Adversarial jailbreaking framework for evaluating LLM safety, with modular experimentation and supervised fine-tuning that improved ethical alignment in Llama-3.2.
Projects
A collection of things I've built
This is the hands-on side of what I do: model evaluation, interpreters, search, pricing tools, mobile apps, and whatever else sounds too interesting to ignore.
Machine Learning & AI
BON Gemma Jailbreak
Recreated Anthropic's Best-of-N jailbreaking work on Gemma-3-1b and ran large-scale adversarial prompt experiments up to N = 5000.
Trained a NanoGPT model from scratch on more than 50,000 pages of Immanuel Kant and used it as a practical way to learn model internals by building one end to end.
DubHacks AI '23 runner-up. A browser extension for voice-command navigation and text-to-speech that pushed accessibility toward a more agentic experience.
Systems Programming
Built a high-performance language VM in Rust and improved interpreter speed with bytecode optimizations and solid unit test coverage.
Husky Hold 'Em
Poker tournament platform for bots, with a Go interface, Python-simulated gameplay logic, and Dockerized participant execution.
Search and recommendation system for University of Washington subreddit posts, with improvements to ranking quality across a few thousand documents.
Other
Options analytics dashboard implementing Black-Scholes, Heston, Jump Diffusion, and Monte Carlo models with visualizations for PnL and Greeks.
Social movie review app for discovery and discussion, built with a native iOS stack and identity management through Auth0.
Mobile app for sustainable outfit generation that used computer vision to help people remix what they already own.
Slate
Production management software for small film productions and creative teams.