Engineered for Data
Privacy & Control.
Built with the security posture enterprise teams expect. Every layer — from transport to application logic — is locked down by default.
Zero Training on Private Data
Audio streams and query content are never used to train, fine-tune, or evaluate any model. Your business data remains exclusively within your infrastructure boundary.
- No audio or transcript retention after session end
- No model training, fine-tuning, or RLHF on customer data
- Fully compatible with SOC 2 Type II data handling requirements
- Session data purged within seconds of disconnection
Role-Based API Sandboxing
The agent operates within strict read-only scopes. Every API call is parameterized and validated against a whitelist — no freeform database access, no arbitrary queries.
- Parameterized queries only — no raw SQL injection surface
- Read-only database connections with row-level scoping
- Whitelist-enforced API endpoints per agent role
- Full audit log of every external call made
Encrypted WebRTC Streams
All voice and data streams use DTLS-SRTP encryption over WebRTC. End-to-end transport layer security is enforced by default with no optional downgrade path.
- DTLS 1.2+ handshake for key exchange
- SRTP encryption for all media streams
- No plaintext fallback — encryption is mandatory
- Certificate pinning prevents MITM attacks
Guardrails & Out-of-Bounds Handling
The Rulebook Agent enforces strict topic boundaries. Any query falling outside the defined policy scope is detected and gracefully deflected — never answered with generated content.
- Intent classifier with confidence thresholds
- Out-of-scope queries receive a deterministic, pre-approved deflection
- No generative fallback for unrecognized intents
- Configurable escalation paths (handoff to human, ticket creation)
How Out-of-Bounds Queries Are Handled
The Rulebook Agent detects off-topic queries and responds with a deterministic, pre-approved deflection — never with generated content.
User Query
"Tell me a joke"
Intent Classifier
off_topic • 0.97 confidence
Scope Check
⛔ Outside defined policies
Graceful Deflection
"I can help with product & policy questions."