Team Idea
PatchPilot is an offline multi-agent system that analyzes a code repository, detects defects, and proposes verified fixes with no human in the loop. Five small, budgeted agents run as a plain Python state machine: Scout maps the repo and ranks suspect files from the issue text and stack traces; Reproducer runs the tests and writes a minimal failing test if none exist; Surgeon generates several candidate unified diffs restricted to the top files; Verifier applies each one in a scratch worktree and rejects anything that breaks a green test; Judge ranks the survivors and emits the fix. Everything runs in a single Docker image on the ROCm base with an open-weights coder model (Qwen2.5-Coder / Qwen3-Coder via vLLM) baked in, so it works with no network. Built solo by Alex Espinoza (UCSB Computer Engineering) with AI coding assistants used openly in development.
Team Members
Unknown Hacker
alexespinoza452810951