OrchevoAn AMI Paradigm platform
Book a discovery workshop
Book a discovery workshop
Reference architecture

Build in the factory. Run as services. Govern through one spine.

An enterprise agentic platform on any perimeter: on-prem, private VPC, hybrid or fully air-gapped — with one identical stack and one governance model.

Enterprise agentic platform

Reference architecture

Factory agents and models promote via GitOps into the runtime, which serves users, apps and events. The Trust Spine applies one set of controls to every action. The platform capability is not a fourth block — it is the substrate every block runs on.

01 · Build · Agentic factory

Agentic Studio + factory assets

Ready-to-use agents, flows and owned models.

Agentic Studio — build & ship

  • Guided & AI builders (no-code)
  • Visual DAG flow designer
  • Pro-code SDK (IDE, notebooks)
  • BYOA: LangGraph / CrewAI adapter
  • Agent manifest · registry
  • Eval harness + red-team gate

Ready-to-use factory assets

  • Pre-built agents (source)
  • Agentic flow templates
  • Model registry & lineage
  • Golden datasets & eval suites
  • Owned model weights — no lock-in
  • Prompt · tool · rail library
02 · Run · Agentic services

Orchestration consuming factory agents & models

Serving users, apps and events.

Consume

  • Copilot: ask · create · execute
  • Embedded apps & channels
  • APIs & SDKs (headless)
  • Admin & tenant console
  • OIDC identity end-to-end

Orchestrate

  • Supervisor & DAG engine
  • Sandboxed agent runtime
  • Memory: run · user · org graph
  • Human-in-the-loop queue
  • Scheduler & event triggers
  • MCP + A2A tool gateway

Serve & retrieve

  • AI gateway (route · rails)
  • vLLM / NIM inference
  • Semantic cache & fallback
  • GovGraph · domain-aware RAG
  • Vector · graph · SQL stores
  • 50+ connectors (CDC, ACL)
03 · Govern · Trust Spine

One set of controls on every action

  • Policy engine (RBAC / ABAC)
  • Approval gates
  • Immutable audit trail
  • Explainability / AI Act
  • OTel traces to the GPU
  • Evals in production
  • Per-agent GPU cost
  • 3-layer guardrails + PII
  • Short-lived secrets vault
  • Confidential computing · signing
Platform capability · deployment agnostic

Not a fourth block — the substrate every block runs on: on-prem · private VPC · hybrid · fully air-gapped.

  • GPU scheduling: priority · MIG · slice
  • Operators & CRDs · GitOps
  • Multi-tenant namespaces
  • Air-gap registry & bundles
Request an architecture briefing See the security model
Deploy it your way

Three deployment patterns. One identical stack and governance.

Only the perimeter changes. Prototype where it is fast, deploy where the data lives.

The flagship

Fully air-gapped

Nothing crosses the wire.

Inside the customer's secure DC — sealed perimeter
  • Orchevo full stack (all components)
  • Models on customer GPUs — local serving
  • Department systems & knowledge
  • Local metering + audit log

Zero telemetry, zero egress. Signed update bundles arrive on approved media, verified and applied in maintenance windows.

In proof of concept: defence forward-location deployment. Usage verified by jointly signed annual audit.

On-premise (connected)

Sovereign estate with controlled egress

The halfway house many estates start from.

Inside the customer's data centre
  • Orchevo full stack (all components)
  • Models on customer GPUs — local serving
  • Department systems & knowledge
  • Controlled gateway for updates

Policy-controlled egress only: update pulls and — where permitted — frontier-model calls through the same governed gateway.

Same governance plane as air-gapped.

Cloud / VPC

Managed speed for eligible workloads

Fast starts, GCC estates, residency-free workloads.

Orchevo-managed cloud / customer VPC
  • Orchevo full stack — managed
  • Frontier & open models via Gateway
  • Connectors to customer systems
  • Same registry, HITL, audit

The migration path: agents built here redeploy into the perimeter unchanged — a starting point, not a fork.

Every meter visible on one dashboard.

One stack, three perimeters: prototype where it is fast, deploy where the data lives — with governance identical on every side of the wire.

Start where the evidence is cheapest

Begin with a two-week discovery workshop.

Three candidate use cases, scored against your data, your perimeter and your stakeholders. Most departments begin here and are in production pilot within two quarters.

Book a discovery workshop See all ways to engage

Two weeks · no cost. You leave with:

  • A use-case shortlist, scored and sequenced
  • A data and perimeter assessment
  • A named executive sponsor identified
  • A scoped proposal you can take to budget