Vibe guardian
15 second video intro: https://www.loom.com/share/3727fd02967d445dab7c8eeb709264e4 1.
Core trick
Injects organization policies and security requirements into coding prompts before generation via a fine-tuned small advisor model.
Description
15 second video intro: https://www.loom.com/share/3727fd02967d445dab7c8eeb709264e4
1. Problem: AI-generated code contains 322% more security vulnerabilities than human-written code (UC Berkeley, Oct 2025), blocking enterprise AI adoption.
2. Solution: Fine-tuned gemini-2.5-flash-lite advisor model that intercepts coding prompts and injects organization policies + security requirements BEFORE code generation.
3. Results: 83% reduction in critical vulnerabilities across 350 test scenarios. 100% elimination of SQL injection and hardcoded secrets.
4. Innovation: Prevention over detection. Based on Berkeley advisor model research (arXiv:2510.02453)—first application to code security.
5. Impact: Real-time (<100ms), works with any LLM, $12 training cost. Enables safe enterprise AI coding adoption.
The linked repo returns 404, so it was deleted or made private.