Writing

1.6M+ Figma instances, no reliable way to know what mapped to real code. I built the infrastructure to measure, classify, and distribute that truth.

Natural-language search over a regulated healthcare directory — without exposing a single record. The agent writes the query; SQL produces the answer.

An automated code reviewer has one currency: whether engineers believe it. A parallel multi-agent review system built to spend every design decision on credibility, not coverage.