Alpie Showcase: Quorum – Offline Adversarial CVE Triage (Team Transformer) #28

🚀 The Big Reveal

• Project Name: Quorum
• The Mastermind(s): Team Transformer

📝 The Elevator Pitch

A CVE lands. Your lockfile says the package is present. The only question that actually matters is: does this vulnerability reach your code?

Quorum is a local-first, adversarial CVE triage system. Instead of trusting a model that claims “94% confidence,” it forces three AI personas (Proposer, Red Team, and Referee) to argue over deterministic facts extracted from the codebase. The Red Team is required to call local MCP tools before it can make any judgment. Confidence is not self-reported — it is measured by how often the panel agrees across multiple seeded runs.

Built for Hacktoberfest Hack Day Hamirpur × GDG Ludhiana (169Pi Track).

🛠️ Alpie Magic & Dogfooding

Quorum is designed as a dual-backend system:

• Primary (local): Gemma models via Ollama for fully offline / air-gapped operation
• Secondary (hosted): 169Pi Playground API using Alpie-Core

The provider is selected at runtime (QUORUM_PROVIDER=auto | ollama | 169pi). When Ollama is unavailable (or when explicitly pinned), the system automatically falls back to Alpie-Core through the Playground API. Only the model client changes — fact gathering, persona logic, seeded rounds, and confidence scoring remain identical on both backends.

We believe Alpie-Core is especially strong for the Proposer and Referee roles thanks to its multi-step reasoning ability, while the tool-calling Red Team can still stay local when needed. We plan to expose backend selection directly in the Streamlit UI in the near future.

Feedback on the hybrid design (local tool-calling + hosted high-reasoning personas) would be very welcome.

🔗 Drop the Links

• Source Code Repo: https://github.com/Team-transformer/quorum-cve-triage
• Live Demo: https://drive.google.com/file/d/1-l8TgSbWaecaAscKoTw16X8dOZPcXb2t/view?usp=drive_link
  (Also includes a Streamlit UI — run with streamlit run app.py. Sample HTML/JSON reports are in the repo.)
• Related Docs: Full pipeline, persona design, confidence math, and MCP tool documentation are available in the /docs folder.