Case Study: Uber's uReview
uReview, Uber's in-house AI code reviewer, analyzes over 90% of the roughly 65,000 diffs landed at Uber each week. Engineers who rate its comments mark 75% useful. Over 65% of its posted comments get addressed in the same changeset. Uber's internal audits put the equivalent figure for human reviewers at 51%. That last comparison is the headline: at Uber, an AI reviewer's comments are acted on more often than a human's.
uReview's design premise is that generation is easy and trust is the hard part. Uber's model evaluation reflects the same split: the best configuration on their golden-comments benchmark paired Claude 4 Sonnet as the comment generator with OpenAI o4-mini-high as the grader. This generator-plus-grader pattern is the same shape we cover in AI reviewing AI.
The mechanism is not a better model. It is relentless filtering, a feedback loop instrumented into every comment, and the discipline to delete whole comment categories that engineers ignore.