Projects

Metric

I built Metric, the recruiting intelligence layer on top of the Cerebral Valley dossiers, with my brother. It is an LLM-assisted evaluation system with pairwise comparisons, where each skill gets its own Elo-style rating instead of collapsing a person into one opaque score.

I worked on the scoring logic, evidence-backed comparisons, and multi-skill evaluations so candidates could be compared by demonstrated strengths with traceable justification. Metric has evaluated 30K+ people and predicted 2 of the top 3 hackathon winners in one tested event. Try to find the easter eggs on the site.

Catan Zero

Catan Zero is a reinforcement-learning project for training models to play Catan. I built it with Michael Tan, who is number one in Catan Elo. We met through YC fellows after he showed me a model trained on scraped Catan games.

Other projects

• SF Graffiti Search: semantic search over 87,790 photos taken by San Francisco city inspectors documenting graffiti violations. Each image was described by Gemini Flash so you can search by content, style, or location. Images sourced by walzr.
• cv-rank (private, email me for more info): debiased applicant comparison with pairwise Elo-style comparisons. It combines LLM-as-judge scoring, rubric scoring, Swiss pairwise rounds, Bradley-Terry fitting, anonymization, position-bias mitigation, debiasing, and borderline review.
• Ranky (private, email me for more info): a human-facing tool for choosing the best photos. It uses pairwise comparison instead of flat scores, so the final order comes from simpler choices with fewer decisions.
• HumanPaste: paste does not work everywhere. It takes clipboard text and retypes it with realistic keystrokes, which is useful in terminals, remote sessions, and apps that accept typing but not direct paste.
• awake: keeps a Mac awake during long AI agent runs, with safeguards so background work does not die when the machine sleeps.
• ultrathink: a drug-discovery search system using AlphaFold, ESMFold, MolGAN, RDKit, and simulated randomness inside an evolutionary search loop.
• DateGuesser-website: weekday calculation trainer built around Conway's Doomsday algorithm, with repeated practice, feedback, and speed modes.
• BoidDynamics: interactive flocking and complex-systems simulation based on Craig Reynolds's Boids algorithm.

Home · Work · About · More on GitHub