GuardianOS

Inspiration I was inspired by Vitalik Buterin's Privacy Pools paper, the SPARC project and the challenge of implementing selective disclosure for blockchain compliance. Current solutions either expose all transaction data or operate without regulatory oversight. I wanted to build a system that only reveals transaction details when absolutely necessary for compliance.

What it does GuardianOS uses multiple AI agents to evaluate blockchain transactions and make compliance decisions. It automatically approves low-risk transactions in under 2 seconds. Medium-risk transactions go through a 5-agent workflow requiring 3-of-5 consensus. High-risk transactions (>€75k or suspicious patterns) trigger my "Tenth Opinion Protocol" - 10 specialized agents working in 4 phases to prevent bias and ensure accuracy. The system only de-anonymizes parties when compliance requires it.

How I built it I used Google ADK v0.5.0 for agent orchestration, creating custom agents extending BaseAgent and SequentialAgent. The backend runs FastAPI on Google Cloud Run with two services - one for compliance, one for fraud monitoring. Smart contracts on Sepolia handle guardian voting. The frontend is Next.js 15.3.3 with WebSocket for real-time updates. MongoDB Atlas provides vector search for pattern matching.

Challenges I ran into The hardest part was coordinating multiple agents without creating bottlenecks. I solved this using ADK's parallel execution for independent analysis and sequential execution for consensus-building. Preventing AI hallucinations in compliance decisions required dedicated quality assurance agents. Synchronizing blockchain events with agent workflows took significant debugging.

Accomplishments that I'm proud of The system implements real threshold cryptography (3-of-5) for guardian consensus. My bias detection catches 7 cognitive bias types. Everything works end-to-end: smart contracts are deployed, APIs are live, and the dashboard updates in real-time.

What I learned Google ADK's structured approach is much cleaner than building custom agent coordination. Separating concerns across specialized agents improves both accuracy and debugging. Threshold cryptography is practical for real-world compliance. Most importantly, privacy and compliance can coexist with intelligent agent design.

What's next for GuardianOS I plan to integrate with actual Privacy Pool implementations when they launch on mainnet. The architecture supports adding new specialized agents without touching core workflows. Long-term, I see this becoming a reference implementation for privacy-preserving compliance systems.

Built With

• fastapi
• google-adk
• nextjs
• python