Engati · AI tooling · company award
AI code reviewer
A reviewer that reads the diff first. Code review went from about 2 hours to 30–60 minutes a day per developer, and the team made it their default.
(01) The problem
Reviewing merge requests took each developer around two hours a day. The idea: let an assistant do the first pass on the diff, while a person still makes the final call.
I designed and built it from scratch, starting at a hackathon; rollout was with the DevOps team.
(02) How it works
- A developer asks for a review from Slack.
- The Spring Boot agent pulls the merge request’s diff from GitLab.
- Token-budget preprocessing shapes the diff to fit the model’s context window.
- An LLM writes review feedback.
- The feedback goes to a human reviewer, who makes the final call.
(03) Key decisions
Start in Slack
Reviews begin where the team already talks, so there was nothing new to learn.
Mind the token budget
Large diffs are shaped before they reach the model, so big merge requests still fit in its context.
An aid, not an autopilot
The AI suggests; a person decides. Its output is always checked by a human.
(04) The hardest tradeoff
LLM calls, not an agent
It started at a hackathon, so I used direct LLM calls on the diff instead of an agent with review skills. That was quick to build and easy to plug into Slack and the CI/CD pipeline, which is how it spread across the company. An agent could reason across more of the codebase, at the cost of a much bigger build.
(05) Results
- Review time per developer per day: ~2 hours → 30–60 minutes. That’s an estimate, based on a senior developer spending about 2 of their 8 hours on reviews.
- Adopted by ~20 developers and became the team’s default review workflow, then spread across the company through Slack and CI/CD.
- Won a company award.