Two AIs review. A third attacks. You get the truth. — ai-flow-architect

[wdnmd1265]

Hi everyone,

I just launched the first version of ai-flow-architect, a multi‑model adversarial code auditing framework.

The problem: Single‑AI code reviewers (like PR‑Agent, CodeRabbit, Copilot) have a 54% false‑positive rate. You trust their output, but they hallucinate.

The solution: Two AIs review your code. A third one tries to break it. You get the truth.

🔍 How it works

1. Brain One – analyzes from a security perspective
2. Opponent – challenges every assumption, using a randomly selected critical style (data‑driven, cost‑driven, edge‑case, etc.)
3. TrustEngine – cross‑verifies both outputs and produces a cryptographically signed TrustReport (SHA‑256 + timestamp)

🎮 Try it live

Playground: https://wdnmd1265.github.io/ai-flow-architect/playground.html

Click any of the three demo cards (SQL injection, logic flaw, contract trap) to see the full 8‑second adversarial review animation, then the final TrustReport.

📦 Install

bash
pip install ai-flow-architect[html]
ai-flow audit ./your-project --html --output report.html

🧠 What makes this different

  | PR‑Agent / CodeRabbit / Copilot | ai‑flow‑architect -- | -- | --

🚀 What’s next

Phase 1 is live (Playground + OG + README). Phase 2 will add:

• Unit/integration tests
• Error‑handling & retries
• Observability (logs/tracing)
• GitHub Action with proper permissions annotation

❓ Feedback wanted

1. Does the Playground animation feel like a “wow” moment?
2. Would you trust a TrustReport with SHA‑256 hashes over a plain‑text review?
3. What’s the biggest blocker for you to try this on a real PR?

---

GitHub repo: https://github.com/wdnmd1265/ai-flow-architect

Thanks for reading, and I’d love to hear your thoughts.