CodeGuardian AI — Gemma 4 on AMD MI300X

CodeGuardian AI is an autonomous software quality engineer that collapses the fragmented code-review workflow — static analyzer → ChatGPT → IDE → commit — into a single transparent AI pipeline. Upload a ZIP of any codebase and watch seven specialized agents run in sequence on a LangGraph StateGraph, each picking up where the last left off: 1. Repository Agent — indexes the codebase into a CodeBundle. 2. Analysis Agent — scans every file with Gemma 4 for bugs, vulnerabilities, and code smells. 3. Prioritization Agent — sorts findings by blast. 4. Explanation Agent — turns each issue into plain English with a recommended fix path. 5. Refactor Agent — proposes a real before/after diff for every issue. 6. Validation Agent — sanity-checks proposed fixes so bad refactors get no credit. 7. Report Agent — composes a Markdown summary with a composite quality score. Every reasoning step — analysis, explanation, refactor, validation — runs on Gemma 4 (26B-A4B-it) served via vLLM on an AMD Instinct MI300X GPU. Token usage is transparent per agent. The whole pipeline finishes in under 90 seconds. Judges can reproduce the run locally with a single `docker compose up --build`. Each issue card in the UI shows a unified red/green diff via react-diff-viewer-continued, and the full report is downloadable as Markdown. Composite score: max(0, 100 - bugs×8 - vulns×12 - smells×3) with per-category caps so the score never collapses to zero on messy codebases. Built for the AMD Developer Hackathon 2026, Unicorn Track, competing for Best Use of Gemma Models.