LinkedIn has concluded that a single AI code reviewer is not enough, so its engineers built a platform that runs several of them at once.
Each reviewer in the lineup uses a different model and reasoning approach. When two or more independently flag the same problem, LinkedIn counts that agreement as strong evidence. Findings raised by only one reviewer still get attention, but they are verified separately before they count. Suggestions that are cosmetic, already fixed, or out of line with the repository are filtered out before anyone sees them.
Under the hood, the system runs on Kubernetes as an event-driven pipeline with durable queues and horizontally scaled workers. The team tracks latency, acceptance rates, completion rates, and provider failures as production metrics, treating review itself as infrastructure.
LinkedIn also built an automated pipeline to measure whether developers actually act on the suggestions. Across 5,230 sampled comments from 1,727 pull requests, 63.9% were accepted. The acceptance rate varied sharply by type: 80% for logic errors, 100% for concurrency bugs, 43.5% for refactoring changes, and 40.6% for security-related fixes.
The design goal is comments that are grounded in the actual diff, tuned to the codebase’s own conventions, and delivered before the human reviewer shows up. Cloudflare and Databricks have published similar efforts, underscoring how central review quality has become as AI-generated code piles up.
