The documents below are shared with technical evaluators, security teams, and compliance reviewers under NDA. Tell us which ones you need and we'll send them over.
How a single AI request becomes a governed action across seven layers: identity, policy, runtime, router, tools, approvals, and audit. What runs inside your environment, and what (if anything) ever leaves it.
Inside: MCP server (SSE) · RS256 JWT identity · per-tenant RBAC + allowlists · single-use TTL approvals · full SHA-256 audit Request the full document →The component-by-component build: agent runtime, cost-optimized router, policy layer, tool registry, and audit trail, plus the supported deployment topologies (private cloud, dedicated VPC, air-gapped) and trust boundaries.
Inside: Docker Compose (on-prem / VPC) · Postgres + pgvector · per-tenant DuckDB · local Ollama · OIDC SSO · pgcrypto secrets Request the full document →How the router dispatches across small specialized models and frontier models (served locally via Ollama or through private endpoints) with policy-driven, confidence- and cost-aware selection and escalation.
Inside: SLM-first via Ollama · confidence-gated escalation · fine-tuned local NL→SQL · local MiniLM embeddings (HF_HUB_OFFLINE) Request the full document →How quality is measured and gated before an agent reaches production: evaluation gates, structured-output validation, versioned prompts and policies, and continuous monitoring after deployment.
Inside: deterministic eval gates in CI · golden sets · Pydantic v2 validation · sqlglot SQL checks · Prometheus metrics Request the full document →