Security, Architecture, and Controls

The full Trust Center — architecture diagrams, data flow, encryption, access controls, and a controls mapping to SOC 2 / ISO 27001 / NIST 800-53 — is being finalised for publication. private·ai is built around three principles: self-hosted, air-gap capable, and audit-native.

• 01Inference runs on customer-managed infrastructure by default — local Ollama serving your chosen model. The provider toggle is itself audited.
• 02Three-layer PII guard pipeline: input guard (blocks PII before the pipeline), content filter (scrubs PII from retrieved chunks), output guard (rejects hallucinated PII patterns).
• 03Persistent audit trail: every prompt, retrieval, generated answer, guard decision, and span is traceable. 90-day Postgres default; Enterprise can archive to immutable / WORM storage indefinitely.
• 04Air-gap capable: container images, model weights, embedding models, and the reranker can all be mirrored to your internal registry. No outbound calls required at runtime.
• 05OpenTelemetry-native observability — every interaction emits a single trace ID that spans the full pipeline, joinable against your existing telemetry.
• 06OIDC / SAML SSO and role-based access control — information-barrier and privilege models configurable per workspace.