Problem
Teams need to know when the AI is confident vs. guessing, keep humans in control of edge cases, and produce audit trails that satisfy compliance and legal.
Solution
A middleware layer that sits between your AI agent and its actions. Every decision gets a confidence score. High-confidence decisions (≥70%) execute automatically and are logged. Low-confidence decisions go to a human review queue before any action fires. Everything is auditable. Inputs, outputs, confidence, latency, outcome, and any human override with reason. A governance report surfaces intervention rates, failure patterns, and performance trends.
Demo
Results
- 3 confidence tiers. Auto-execute, human review, escalate
- Full audit trail on every AI decision
- Human override with reason logged
- Governance report with explainability summary
- Live demo. HR candidate screening scenario
Every incoming request (a candidate summary, a document, a ticket) is sent to Claude via OpenRouter for classification. Claude returns structured JSON with type, priority, routing recommendation, and a confidence score. AgentGuard's routing engine evaluates the confidence: ≥70% triggers automatic execution and audit logging; <70% queues the decision for a human reviewer who can approve, reject, or override with a reason. All outcomes write to a SQLite audit log with full input/output context, latency, and model used. The /report endpoint computes intervention rates, average confidence by category, failure patterns, and a plain-English explainability summary. Frontend is vanilla HTML/JS with a dark theme: intentionally no framework so the governance logic is the focus.
Tech Stack
- Node.js
- Express
- Claude AI
- OpenRouter
- SQLite
- Vanilla HTML/CSS/JS
- Vercel