High-Signal AI Code Review That Adapts to Your Codebase at Scale

AI coding tools are generating more code, faster, but human code review remains a bottleneck. At LinkedIn, where tens of thousands of pull requests are opened each week across roughly 10,000 repositories, we asked: Can AI keep up with AI without overwhelming developers with noise?

In this talk, I’ll share how we turned an early AI code-review experiment into infrastructure that engineers now rely on. The system combines multiple independent review agents to reduce model blind spots, a three-tier customization framework to apply organization-, repository-, and use-case-specific knowledge, and a validation pipeline that filters low-confidence and redundant findings before they reach developers.

I’ll also discuss the production lessons that mattered beyond model selection: ensuring reviews arrive before humans begin reviewing, evaluating every model and prompt change before release, operating reliably across providers, and measuring whether suggestions are reflected in the final merged code rather than relying on reactions or surveys.

Today, the system reviews tens of thousands of pull requests each week, usually within minutes, with more than 60% of sampled suggestions accepted. Attendees will leave with practical architecture and evaluation patterns for building AI code-review systems that are timely, adaptable, measurable, and most importantly, useful to human reviewers.