secure-edge-ai-governance

Why this project exists

Edge AI governance is often documented as policy, diagrams, or compliance checklists. Those artifacts matter, but they do not prove that a release decision is actually enforced.

This project turns governance into executable software. A candidate release is approved only when deployment identity, two distinct human approvals, risk, drift, signature, attestation, and regression evidence satisfy deterministic policy. Any missing or failed control produces manual_hold.

Advisory AI can recommend. Deterministic policy and accountable humans authorize.

Related MCP server: @vaibot/mcp-server

Core engineering principle

The system deliberately separates:

• probabilistic AI advice from release authority;

• human accountability from automated scoring;

• evidence collection from policy evaluation;

• model quality from device trust; and

• demo claims from production security capabilities.

The objective is not simply to predict whether a release is safe. The objective is to make the release decision inspectable, testable, fail-closed, and reproducible.

Release gate

flowchart TD
    A[Deployment candidate] --> B[Deployment identity]
    B --> C[Two distinct approvers]
    C --> D[Risk within limit]
    D --> E[Drift within limit]
    E --> F[Signature valid]
    F --> G[Attestation valid]
    G --> H[Regression passed]
    H -->|all pass| I[APPROVED]
    B -->|fail| X[MANUAL HOLD]
    C -->|fail| X
    D -->|fail| X
    E -->|fail| X
    F -->|fail| X
    G -->|fail| X
    H -->|fail| X

Architecture

flowchart LR
    A[Model / Firmware / Wasm Candidate] --> B[Cloud Control Plane]
    H[Two Human Approvers] --> B
    AI[Advisory AI\nRisk · Drift · Optimization] --> B
    B --> G{Deterministic Governance Gate}
    G -->|Approved| S[Sign / Stage Release]
    G -->|Failed evidence| M[Manual Hold]
    S --> T[Secure Transport + Attestation]
    T --> D[Edge Device Runtime]
    D --> O[Health / Drift / Telemetry]
    O --> B
    O -->|Failure| R[Known-good Rollback]
    R --> D

What is implemented

Verify it in two minutes

1. Open the live Streamlit demo.

2. Keep all evidence valid and use two distinct approvers → approved.

3. Change risk above the threshold, invalidate attestation, fail regression, reuse one approver identity, or remove an approval → manual_hold.

4. Inspect lib/policy.ts and demos/streamlit/policy.py to see the policy implemented independently in TypeScript and Python.

5. Check GitHub Actions for repeatable verification.

MCP governance tools

The /api/mcp JSON-RPC surface exposes bounded governance operations:

An AI agent may call these tools, but it cannot override the deterministic policy implemented inside them.

Security boundary

Implemented as executable controls:

• deterministic fail-closed decision logic;

• two-person approval semantics;

• risk and drift limits;

• signature, attestation, and regression evidence checks;

• explicit hold reasons; and

• independent CI verification.

Not represented as completed production infrastructure:

• identity-provider authorization;

• hardware-backed TPM/TEE quote verification;

• HSM/KMS production release signing;

• production mTLS lifecycle;

• append-only audit storage;

• real OTA fleet orchestration; and

• production rollback execution.

See docs/SECURITY_MODEL.md for the hardening path.

Run locally

git clone https://github.com/h00w/secure-edge-ai-governance.git
cd secure-edge-ai-governance
npm ci
npm run lint
npm test
npm run dev

Streamlit reviewer demo:

cd demos/streamlit
python -m pip install -r requirements.txt
python -m pytest -q
python -m streamlit run app.py

Proof chain

Architecture → Policy → Interactive UI → MCP Tool Surface → Tests → CI → Documentation → Live Demo

• Live demo: https://secure-edge-ai-governance.streamlit.app/

• Architecture: docs/ARCHITECTURE.md

• Security model: docs/SECURITY_MODEL.md

• Demo guide: docs/DEMO_GUIDE.md

Author

Hendarmawan, PhD Eng.
AI Engineering · Secure Edge AI · Trusted Computing · AI Governance · Production AI

Website · LinkedIn · GitHub