AI-assisted coding gave us a good problem at Lendsqr: engineers were pushing more PRs than ever. Writing code had become faster. Reviewing it still required someone to read the change, understand the problem, and make a judgment. That someone was often me, while my own tickets sat in the corner looking disappointed.
“Jemimah hasn’t reviewed my PR” became a regular feature of engineering updates. Some of those timelines were a little creative, but I digress. 😂
The bottleneck was real. More code was reaching review, but the number of hours in my day had not received a corresponding upgrade. So I built Revisar.
My Mac went to sleep
The first version was wonderfully simple: the GitHub CLI fetched PRs, and Codex reviewed them. It worked right up until my Mac went to sleep. A reviewer that needed me to keep my laptop awake was not going to fix an availability problem, so I moved it to an Oracle server and kept developing it into a service that runs independently of my machine.
Revisar now checks the repositories every 10 minutes, waits for CI to pass, and posts reviews directly on GitHub. It also checks ClickUp for a real task behind the PR. That is not paperwork for its own sake: a review needs to know what the change is supposed to fix before it can say whether the code actually fixes it.
The worker keeps a persistent queue and reviews the current change, not a stale copy of yesterday’s PR. If an engineer pushes a new patch, that patch needs a new look. The feedback stays in GitHub, where the conversation about the code is already happening.
A review needs the ticket, not just the diff
Getting AI to comment on a diff was the easy part. A useful review has to follow affected code paths, look for regressions and performance problems, check our engineering standards, and compare the implementation with the task that prompted it. Revisar verifies its proposed findings before posting them. When it finds a real issue, it should explain what is wrong, why it matters, and how to fix it, ideally in one coherent review instead of a trail of vague comments.
This is also where the tool keeps improving. When it misses something I catch, I update its review standards and test the same kind of issue again. It is less glamorous than naming an AI colleague, but it is how the reviews become more consistent.
The robot does not approve money changes
Revisar has a confidence rating. I was still not about to let a robot merge code just because it felt confident, especially a robot I built myself. 😭
It can move a narrow set of low-risk fixes into development or staging when the checks and live PR state still allow it. Core functionality, financial operations, uncertain changes, and production releases stay with a human reviewer. A green build is useful evidence; it is not a license to move money-changing code unattended. When Revisar cannot establish the answer, it explains what a person needs to decide.
That boundary is part of the product. Engineers can get a first round of actionable feedback in roughly 10–15 minutes when the checks are ready, often before I have opened my laptop. Small issues can be resolved early, leaving human review for the decisions that actually need it. The PR moves through a clearer flow, and I get to work on my own tickets too. A minor miracle.
This is the kind of product engineering I enjoy: finding the bottleneck that slows down the people building customer-facing features, then building a tool that makes their work better. We used AI to make writing code faster. I built Revisar because reviewing it needed to catch up.
Meet my little robot colleague. It does not sleep, it does not complain, and thankfully it does not have tickets of its own. At least now, when someone says “Jemimah hasn’t reviewed my PR,” I have a robot to defend myself. 😂