Designing trust into AI code review

Making an AI reviewer's findings legible and actionable, so engineers act on them instead of muting the bot.

The problem

An AI reviewer that comments on every pull request is only useful if developers trust it. Too noisy and they mute it; too quiet and they ignore it. Trust is the whole product.

What I designed

A review experience built around confidence and provenance — every finding shows why it fired, how sure the model is, and what to do next, with one-click dismiss-and-learn.

Severity that means something. Findings are ranked so the important ones surface.
Show your work. Each comment links to the exact lines and the reasoning behind it.
Teach the bot. Dismissals feed back into what it flags next time.

Outcome

Higher action-rate on comments and lower mute-rate — the signals that an automated reviewer is actually trusted. (Metrics directional and confidential.)

This second case study exists to demonstrate scalability: it’s just another .md file in src/content/projects/. The listing card and this page were generated automatically.