Skip to content
← All work

P-007 · 2026

PRPilot — self-hosted AI code review

An open-source GitHub App that turns pull requests into queued, multi-model code reviews while keeping source code, findings, and audit history on infrastructure teams control.

Role
Founder · solo engineer
Year
2026
Status
Open source
Outcome
Self-hosted · multi-model · MIT
GitHub repository
Next.js 16React 19FastAPIKafkaCeleryRedisPostgreSQLLiteLLMGitHub App / APIDocker / Kubernetes

The problem

AI code review is useful only when it is dependable, explainable, and safe to run against real repositories. A synchronous webhook handler cannot absorb model latency or provider failures, while a hosted reviewer can make source code and findings leave the team's infrastructure. PRPilot separates ingestion, queueing, model execution, persistence, and GitHub delivery so every review can be retried, audited, and self-hosted.

Architecture

Interactive system view

See the product from purpose to implementation.

Follow the work, then switch views when you want the practical answer: who it is for and what it is built with.

Flow pausedStep 01 / 05

Step 01 of 05

A GitHub App receives pull-request…

A GitHub App receives pull-request events at a FastAPI webhook receiver. HMAC verification rejects forged payloads before any repository content enters the system, and Redis-backed idempotency prevents duplicate reviews when GitHub retries delivery.

Full system map

The grouped dependency view stays visible so the infrastructure is readable at a glance.

System map

How the pieces connect

PRPilot is intentionally a pipeline: GitHub sends a signed event, FastAPI verifies and deduplicates it, Kafka makes the work durable, and Celery performs the review outside the webhook request. LiteLLM keeps model providers swappable; PostgreSQL stores the durable review and audit record; Redis handles locks and idempotency; the dashboard observes the system; and the GitHub API closes the loop by publishing the result back to the pull request.

Outcome

  • Open-source, self-hostable GitHub App with MIT licensing.
  • Durable Kafka → Celery processing keeps webhook ingestion independent from model latency.
  • Multi-provider review routing through LiteLLM, including local Ollama deployments.
  • Dashboard, audit history, Redis idempotency, and PostgreSQL persistence around every review.

Lessons learned

  • Webhook ingestion, model execution, and GitHub delivery should be separate failure domains.
  • Idempotency belongs at the boundary because retries are normal behaviour, not exceptional behaviour.
  • A model gateway is most valuable when it preserves product policy while providers remain replaceable.