Autonomous forensic auditor for AI voice agents.
Extracts governance rules, scores call transcripts, generates prompt fixes.
AI voice agents are hard to audit — they produce transcripts, but nothing automatically checks whether the agent followed its own instructions. This tool closes that gap.
Give it a Vapi voice agent. It:
- Analyzes the system prompt and extracts a governance blueprint — the rules the agent is supposed to follow
- Scores real call transcripts against that blueprint, flagging every violation
- Generates developer-ready prompt fixes and a management-level report
The result: a full forensic audit of your voice agent's behaviour, end-to-end, with no human review required.
Vapi voice agent (system prompt + call transcripts)
│
▼
① Analyzer agent
Extracts governance blueprint from system prompt
│
▼
② Scorer agent
Audits each transcript against the blueprint
Flags violations with severity + context
│
▼
③ Optimizer agent
Generates prompt rewrites for each violation
Produces a management summary report
| Layer | Role |
|---|---|
| OpenClaw | Core multi-agent framework — orchestrates the three sub-agents |
| Vapi | Voice agent source — pulls system prompts and call transcripts |
| n8n | Action and alerting layer — routes reports and triggers |
Built as a hackathon proof-of-concept. The core insight: voice agent governance is a retrieval + reasoning problem — you have a spec (the system prompt) and evidence (the transcripts), and an LLM can act as a reliable judge between the two.
The three-agent pipeline (Analyze → Score → Optimize) keeps each step focused and auditable. OpenClaw provides the multi-agent orchestration layer so the agents can run in parallel across large transcript volumes.
MIT