All projects

Sentinel Orchestrator Network

AI agents that check a Cardano policy for fraud, with a verdict, a PDF report and a governance module for DReps.

Kind
Multi-agent system for Cardano
Year
2025
Built with
  • Cardano
  • Hydra
  • Blockfrost
  • Koios
  • Python
  • FastAPI

The question

Before you sign a Cardano transaction or hold a token, how can you tell it is not a scam?

Sentinel asks a small crew of agents to inspect the policy and gives you one verdict with the reasons.

Why it exists

Cardano’s DeFi ecosystem is growing without real-time threat detection or trustless oracle infrastructure. Users and protocols rely on separate security tools, centralised data feeds and manual checks. Built for a Cardano hackathon.

How it works

From a policy ID to a verdict

  1. 1

    Scan

    A policy ID or transaction is submitted and queued as a background task.

  2. 2

    Compliance

    The Sentinel agent checks its format and a list of known-bad patterns, and stops here for an invalid input.

  3. 3

    Specialists

    The Oracle runs five checks at once, each isolated so that one failure does not stop the others.

    Block scanner
    Forks and chain tip.
    Stake analyser and vote doctor
    Stake and voting patterns.
    Mempool sniffer and replay detector
    Pending activity and repeated transaction patterns.
  4. 4

    Fusion

    A weighted average of the risk scores, raised to a floor for critical and high findings.

  5. 5

    Verdict

    Safe, warning or danger, with a PDF report.

Decisions

  • Specialists run in parallel and fail alone

    Each specialist gets its own timeout. If one fails it is counted as a low-risk result marked unsuccessful, and the others still report.

  • Severity overrides the average

    A critical finding lifts the risk to at least 0.95 and a high one to at least 0.75, so a single serious signal cannot be averaged away.

  • A fast path through Hydra

    When a Hydra head is available, a transaction is validated there first and the standard flow runs only if that path fails.

Evidence

  • Built

    The governance module is the largest part of the agent code, with about 1,380 lines across five modules.

Not built yet

  • The Compliance, ZK-Prover and Consensus agents of the five-agent design
  • Midnight zero-knowledge proofs and minting a ThreatProof on Cardano

Built with

Languages

  • Python

Frameworks and libraries

  • Blockfrost
  • Koios
  • FastAPI

Protocols and chains

  • Cardano
  • Hydra