Inspiration

Code reviews are slow. Developers wait hours for feedback, bugs slip through, and junior devs get no mentorship. I wanted an AI agent that reviews code instantly, catches security flaws, and generates tests automatically.

What it does

DuoReview analyzes merge requests in 30 seconds and:

  • Detects bugs, security vulnerabilities, and anti-patterns
  • Generates missing unit tests automatically
  • Posts detailed comments with fix suggestions
  • Gives a quality score out of 10
  • Provides a clear verdict: Approved, Changes Requested, or Blocked

How we built it

Built on GitLab Duo Agent Platform using:

  • Claude Sonnet 4.5 for AI analysis
  • YAML configuration for agent and flow
  • 22 GitLab tools (read files, analyze code, create tests, post comments)
  • Structured 6-step review process

Two modes:

  1. Agent (chat): Invoke manually in GitLab Duo Chat
  2. Flow (automated): Triggers on new merge requests

Challenges we ran into

  • Learning the Agent Platform from scratch in a few days
  • Service account permissions were tricky to configure
  • Balancing the scoring algorithm (not too harsh, not too lenient)
  • Making generated tests follow each project's framework conventions
  • Handling large MRs without overwhelming feedback

Accomplishments that we're proud of

  • Built a fully functional code review agent in less than a week
  • Detects 20+ issues in 30 seconds vs 45 minutes for humans
  • 95% security vulnerability detection rate
  • Auto-generates 30-40 comprehensive tests per file
  • Works across 8+ programming languages

What we learned

  • Prompt engineering is 80% of building a good agent
  • AI needs structure—the 6-step process makes reviews consistent
  • Tools define what agents can do (safety through constraints)
  • Even experienced devs miss SQL injections and hardcoded secrets
  • AI complements humans, doesn't replace them

What's next for DuoReview

  • Add support for more languages (Kotlin, Swift, Scala)
  • Integrate with CI/CD for automated quality gates
  • Build a dashboard to track code quality trends over time
  • Add custom rule configuration per project
  • Implement learning from past reviews to improve detection

Built With

Share this project:

Updates