AI Powered Merge Request Review with Gemini AI

An AI assisted pull request reviewer to reduce manual efforts and provide instant feedback along with suggestions and high-risk classification.

Comment

Inspiration

Code reviews are essential, but they often slow down development. In many teams, even small changes can sit unreviewed for hours or days. I wanted to explore whether AI could act as a first-pass reviewer — something that gives immediate feedback and reduces the initial review burden.

The goal wasn’t to replace human reviewers, but to add a lightweight layer of automation that improves developer productivity.

What it does

This project is an AI-powered pull request reviewer for GitLab.

Whenever a merge request is created or updated:

  • A GitLab CI pipeline is triggered automatically
  • The pipeline runs a Python script that fetches the code changes
  • The changes are sent to Gemini AI for analysis
  • The AI generates structured feedback (issues, improvements, and risk level)
  • The feedback is posted directly on the merge request as a comment

Additionally, the system can apply labels (like high risk) based on the AI’s assessment.

How we built it

The solution is built using:

  • GitLab CI/CD for automation
  • Python for scripting
  • GitLab REST APIs to fetch merge request data and post comments
  • Gemini API for AI-based code analysis

The core logic lives in a single script (review.py) that:

  1. Reads merge request details from environment variables
  2. Fetches the diff using GitLab APIs
  3. Sends a structured prompt to Gemini
  4. Parses the response
  5. Posts the result back to the merge request

The pipeline is defined in .gitlab-ci.yml and is triggered on merge request events.

Challenges we ran into

  • API limitations and model compatibility
    Initially tried OpenAI but ran into quota issues. Switching to Gemini required figuring out the correct model names and endpoints.

  • Managing input size for AI
    Full diffs can be large, so we had to limit the number of files and truncate diffs to stay within token limits.

  • Ensuring structured AI output
    Getting consistent and usable responses required refining the prompt multiple times.

Accomplishments that we're proud of

  • Built a working end-to-end AI reviewer integrated directly into GitLab
  • Automated the full flow from merge request → analysis → feedback → action
  • Added risk-based labeling to make the system actionable, not just informative
  • Kept the implementation simple while still demonstrating real value

What we learned

  • CI/CD pipelines can be used for much more than builds and deployments
  • AI is easy to integrate, but making it reliable and structured takes effort
  • Small design decisions (like prompt format and filtering) significantly impact output quality
  • Understanding how tools like GitLab actually execute pipelines is critical for debugging

What's next for AI Powered Merge Request Review with Gemini AI

  • Add inline comments on specific lines of code instead of a single summary
  • Improve diff handling to include more context intelligently
  • Add security and best-practice checks
  • Make model selection configurable via environment variables
  • Explore multi-agent workflows for deeper analysis

Built With

  • ci/cd
  • devops
  • gitlab-api
  • gitlab-ci/cd
  • google-gemini-api
  • pipelines
  • python
  • requests-library
  • rest-apis
Share this project:

Updates