cat ./projects/raptorreview-ai.md
[PUBLISHED · GITHUB MARKETPLACE]02 /
RaptorReview AI
AI code review, straight from the GitHub Marketplace.

// stack
- Python
- GitHub Actions
- GitHub API
- Groq
- tiktoken
- YAML
// THE PROBLEM
PR reviews under deadline pressure get skimmed. Security issues, N+1 queries, and hardcoded secrets slip through not because developers are careless but because reading a 400-line diff for subtle issues is cognitively exhausting. Existing AI review tools require a paid API key, a hosted service, or both — too much friction for individual developers or small teams.
// WHAT I BUILT
A GitHub Action that automatically reviews every pull request using Groq's LLM inference — posting line-specific security, performance, and readability comments directly in the PR diff interface. Built with Python and tiktoken for context window management. The action reads the PR diff via the GitHub API, segments it into reviewable hunks, and posts inline comments on specific changed lines. No hosted infrastructure. No paid API key. Installable by any GitHub repository in one step via the Marketplace listing.
// THE KEY DECISION
Review only the changed lines (diff hunks), not the full file context. This keeps the token count per PR proportional to the change size, not the file size — a 10-line fix in a 2,000-line file sends 10 lines to the LLM, not 2,000. tiktoken enforces the context window limit before the API call, so the action never fails due to an oversized prompt. This was the core engineering constraint that made the tool usable on real codebases.
// WHAT IT TAUGHT ME
Building for the GitHub Actions runtime is different from building a web app. Your code runs in a fresh container on every PR — you control nothing about the execution environment. The action.yml interface (inputs, environment variables, entrypoint) is the entire API contract between the user and the tool. Getting the Marketplace listing right required understanding GitHub’s verification process and the action metadata spec in detail.
// REAL-WORLD IMPACT
Published on GitHub Marketplace — installable by any developer or team in a single workflow YAML addition. An automated security and code quality pass on every PR, at zero cost. The tool catches what tired human reviewers miss: hardcoded credentials, SQL injection patterns, missing null checks, and performance anti-patterns in the changed lines.