Daksh Gupta, co-founder of Greptile, explains how to build or assess an automated code reviewer in this eighteen-minute AI Engineer talk.

Start simple, then measure what breaks

  • Connect a pull-request webhook to an agent in a sandbox and return findings through the repository’s review API.
  • Expect missed bugs, missing context, expensive runs, repeated findings and technically correct but distracting nitpicks.
  • Gupta reports seeing internal reviews cost $3–25 per pull request; he provides no standardized workload for that range.
  • Target bug types from past incidents and explicitly define which low-value comments to suppress.

Build evidence before tuning

  • Assemble historical pull requests that caused incidents and check whether each reviewer version catches the known defects.
  • Refresh that dataset as coding agents introduce new failure patterns.
  • Try a different model family for review than the one that wrote the code; Gupta reports a 10% higher catch rate in Greptile’s experience.
  • Fewer comments alone cannot tell you whether noise fell or important bugs disappeared from the review.

Give the reviewer code and intent

  • Index subsystem behavior, relevant files and flows so the agent can retrieve precise codebase context.
  • Include documentation and tickets to check whether working code fulfills the requested change.
  • Gupta describes a living wiki combining code, product plans, logs and production information.
  • Run the application in a dependency-ready sandbox and test documented user flows, including device-specific rendering.

Spend stronger models where they matter

  • Route context selection and routine exploration to cheaper models; reserve stronger models for difficult bug analysis.
  • Gupta says Greptile uses roughly 20 task types with task-specific model choices; validate substitutions on the incident dataset.
  • Caveat: the catch-rate and cost claims are vendor observations. The talk supplies no reproducible comparison or evidence that automated review can safely replace human release judgment.

“Your correct code can actually still be quite wrong.” — Daksh Gupta, 13:13–13:15