cd /news/ai-tools/i-built-an-ai-code-review-github-app… · home topics ai-tools article
[ARTICLE · art-24861] src=dev.to pub= topic=ai-tools verified=true sentiment=↑ positive

I built an AI code review GitHub App that tracks your team's mistake patterns, here's how

Ahmad Mustafa built CodePulse, a GitHub App that analyzes pull requests and tracks recurring code issues per developer over time. The app runs a two-pass AI analysis using Groq's Llama 3.3 70B model, stores findings per developer, and sends personalized weekly email digests summarizing each person's recurring mistake patterns. CodePulse is live at getcodepulse.vercel.app and open-source on GitHub.

read2 min publishedJun 12, 2026

Most AI code review tools tell you what's wrong in a PR. CodePulse tells you what your team keeps getting wrong, week after week.

I built CodePulse as a GitHub App that does two things regular code review bots don't: it stores every finding per developer over time, and sends a personalized weekly digest summarizing each person's recurring issue categories. The idea is simple — if you're writing the same type of bug every week, you should know about it.

How it works

Install the GitHub App on your repos. Every time a PR is opened, CodePulse fetches the diff, runs a two-pass AI analysis using Groq's Llama 3.3 70B with structured tool-calling, and posts inline review comments pinned to exact lines with severity labels — Critical, High, Medium, Low.

Every finding gets stored against the developer who wrote it, the repo, and the PR. This is what makes CodePulse different from a one-shot reviewer — it builds a history.

Every Sunday, developers who opt in get a personalized email digest aggregating their issue categories from the past week. Not a wall of comments — a summary of patterns.

The Stack

Backend runs on Node.js, Express, TypeScript with Prisma ORM and Neon PostgreSQL on Azure. Frontend is React with TanStack Router on Vercel. Weekly digests go out via Resend, triggered by a GitHub Actions cron job.

The trickiest part was the two-pass analysis — a file triage pass first to skip lockfiles, minified assets, and generated files, then a chunked deep review pass that returns typed JSON per issue via structured tool-calling. Getting Groq to return consistent structured output at speed was the core engineering challenge.

What I learned

Webhook reliability is everything in a GitHub App. HMAC-SHA256 verification, idempotent processing, and logging every delivery from GitHub's Advanced tab saved me hours of debugging. If you're building on GitHub webhooks, treat the Recent Deliveries panel as your best friend.

Multi-tenant isolation also required more upfront schema design than I expected, scoping every query by GitHub App installation ID from day one is non-negotiable.

Try it

Live at getcodepulse.vercel.app — install the GitHub App on any repo and open a PR. Inline comments appear within 1–3 minutes.

Source on GitHub: github.com/ahmadmustafa02/CodePulse

I'm Ahmad Mustafa, a Full Stack Developer based in Islamabad, Pakistan. Building AI-powered products and publishing research in deep learning for IoT security and medical imaging.

── more in #ai-tools 4 stories · sorted by recency
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
Live at https://your-agent.zahid.host
Get free account → Pricing
from €0/mo · no card required
LIVE [news/i-built-an-ai-code-r…] indexed:0 read:2min 2026-06-12 ·