AI systems
Code Buddy
Code review at engineering scale. I architected a multi-agent platform at Qualcomm that finds issues, proposes fixes, and validates them across 4,000+ repositories.
CODE BUDDY01 / AGENT SYSTEMS
review.pipeline
01// From a change to a verified fix
02const review = await agents.run({
03detect: issues,
04verify: tests + guardrails,
05deliver: actionable fixes
06});
From detection to a self-validating fix
12,000+ engineers4,000+ repositories
About the project
Code Buddy separates issue detection from repair. Agents retrieve symbols and dependencies, generate candidate fixes, and validate changes before posting a review. I architected the platform at Qualcomm, where it reached 12,000+ engineers, 4,000+ repositories, and more than two million automated reviews.
What I built
- Architected a multi-agent GenAI code-review platform deployed across 4,000+ repos for 12,000+ engineers , running detect-then-fix agents under a tool-calling orchestrator that generate self-validating fixes and pass validation guardrails before posting.
- Scaled the platform to 2M+ automated code reviews (180,000+ monthly), with a modeled estimate of 114,000+ engineer-hours saved per month .
- Cut review-retrieval latency via a Tree-sitter symbolic-retrieval layer over a ccls/ctags dependency graph with TTL symbol caching.
- Built SlideGen , an LLM tool that turns raw technical data into corporate-compliant decks using structured-output validation and template-aware layout.
- Built an agentic tool that reads live C/C++ modem source and drives interactive 4G/5G NAS protocol tests through a 14-tool calling loop, no hardcoded answer mappings.
- Built an autonomous LLM agent that triages NAS modem hotline escalations end-to-end via a 13-tool calling loop: parses issue emails, searches Perforce source and CR/JIRA databases, and shelves a 3GPP-cited candidate fix after a 4-model consensus review.
- Built evaluation workflows for code-review agents using golden code changes and precision/recall comparisons. Worked on structural code retrieval so agents could look up symbols and dependencies instead of relying on lexical search alone.