ADI Platform · Security
Guardrails in code, verified outside the model
Three mandatory filters, eight configurable, three severity levels and an auditable per-turn trail.
An instruction in a prompt is not a control
Asking the model not to say something is a preference, not a guarantee. A Risk team does not approve a control the controlled system itself can ignore.
So the filters run before and after the LLM, in deterministic code, and the model does not pick the final action: it proposes within a closed catalogue whose enums are validated outside.
The filters
Permissive, moderate and strict. The three mandatory filters cannot be disabled on any agent.
What Risk can reconstruct
Per-turn trail
What the agent said, on which prompt and knowledge base version, and which filter acted.
PII tokenised at the edge
Personal data never reaches the model provider; the audit log records the interaction without the data.
Versioned regression QA
Every behaviour change is tested against the approved case set.
Canary and rollback
Gradual percentage release, with a one-step return to the previous version.
Per-organisation isolation
Multi-tenancy with server-side enforcement, not client-side.
Deployment in your perimeter
Dedicated VPC or private network, whenever policy requires it.