Making “this doesn’t sound right” testable
I found the failure modes, wrote the assertions and a response-quality rubric, and calibrated an LLM judge against human review.
Selected work
2025-2026What good looks like (opens in a new tab)
A live tool that puts the evals method to work: choose the standards, apply them, watch an LLM judge grade the lift.
One vocabulary, three surfaces
A controlled vocabulary and in-product messaging, written for people and encoded as an agent-readable skill.
What AI knows about you
UI and labels that frame AI insights and recommendations, with assertions to measure the output.
Push notification framework
A more relevant, engaging push-notification system, grounded in user research and shipped as updates.
AI tradeoff conversations
Reading user intent, naming a tradeoff, and recovering from a bad guess.
Prompts for personalized recommendations
Iterating a system prompt and updating a Python script to lift output quality.
My case studies are password-protected. If you’re a recruiter or hiring manager, reach out and I’ll walk you through them.