AI Advisory Plane

Three role-specific agents on one governed AI Advisory Plane

Vannarho RaaS runs Config Doctor, Control Plane, and Explainable Results agents on one governed architecture aligned to the run lifecycle, not three separate AI stacks.

Tenant-scoped context Policy-checked invocation Cited and auditable output

1. Config Doctor Agent

Onboarding and pre-execution readiness

Customer experience: during onboarding and before launching an analytic run, the user gets a clear go/no-go view with specific issues, required client inputs, and recommended repair actions.

What it does before launch

  • Determines if a submission is runnable
  • Evaluates bind readiness, scope/config compatibility, missing inputs, waivers/approvals, unsupported seams
  • Produces case, findings/reports, client questions, repair plans, predictive hints, evidence bundle
  • Finalizes gate outcome as RUNNABLE, RUNNABLE_WITH_APPROVED_WARNINGS, or blocked outcomes

Governance boundary

  • Advisory-only for governance decisions; can draft fixes but cannot self-approve
  • No run request allowed unless Config Doctor is FINALISED with launch_allowed=true

Retrieval defaults: elaborated embeddings, known good/bad configs, and hashed-tfidf retrieval in core Config Doctor service logic.

Advisor CLI: sentence-transformers with --embedding-model all-MiniLM-L6-v2.

LLM task default: gemini.

2. Control Plane Agent

Run health and operational interpretation

Customer experience: during incidents or delays, users get a plain-language explanation of what happened, where it failed, and the safest next action.

Scope and interpretation role

  • Covers execution and blocked-serving phases (Run Health Analyst)
  • Interprets dispatch, job, load, alert, and serving lifecycle state
  • Drafts safe operational next steps and support actions

Authority and limits

  • Uses typed control-plane and evidence records as authority
  • Read-plus-draft only
  • No bypass of approvals or typed workflows

3. Explainable Results Agent

Post-execution result interpretation

Customer experience: after results are loaded, users can ask what changed and why and receive cited, regulator-safe explanations linked to evidence.

Signals and narrative sources

  • Uses governed BigQuery facts, serving/comparison views, and evidence references
  • Uses BigQuery ML for deterministic signals
  • Uses Vertex AI for bounded narrative and Q/A

Deterministic truth remains fixed

  • Explains, compares, and drafts commentary
  • Does not alter deterministic result truth

Shared AI Advisory Plane

One governed model behavior across all three roles

Customer experience: consistent assistant behavior across readiness, operations, and results, instead of three disconnected AI experiences.

  • Tenant-scoped, policy-checked, typed context bundles before model invocation
  • Cited, replayable, auditable outputs
  • Real mutations/publication routed only through authorized APIs and approvals

Google-native architecture mapping

Advisory plane integrated with production services

Apigee/API Gateway + control plane APIs, Cloud Tasks, GKE (Axion) execution with ADK, GCS evidence custody, BigQuery canonical/serving/BI layers, Dataflow loaders, Firestore hot projections, and Vertex AI reasoning.

Apigee/API Gateway Cloud Tasks GKE + ADK GCS evidence BigQuery + BI Dataflow Firestore Vertex AI

Governed AI agents, deterministic risk truth

The advisory plane can diagnose, interpret, and explain. Deterministic pricing, approvals, and publication remain in governed workflows.

See AI boundary Help shape roadmap