Project case study
CodeReview Bot
AI code review that catches hygiene issues so senior engineers can focus on architecture. Claude via Bedrock, chunked analysis, and per-repo rate limiting.
Problem
- Senior engineers were spending review time on pattern violations and style issues
- Architecture and logic, the stuff that actually needs human judgement, was getting squeezed
Constraints
- Handles 15+ languages
- Rate limiting to keep API costs predictable
- Feedback has to be specific and actionable, not generic lint warnings
- Humans still own architecture decisions
Architecture
- Frontend shows syntax-highlighted diffs
- Backend splits code into context-window-safe chunks
- Language-specific prompts go to the LLM, feedback is aggregated back
- Rate limits applied per user and per repo
Tradeoffs
- Claude over GPT-4 at the time. Gave more specific code-review feedback on my test set
- Chunking over full-file analysis to stay inside context windows reliably
- App-level rate limiting over raw throttling. Better user experience and predictable cost
Outcome
- Routine review comments are now automated
- Senior engineer review time shifted to architecture and design