A HEPO deployment places the application services, working data, files and local inference within an agreed infrastructure boundary. The organisation and implementer establish where those components run, who operates them and how they connect to existing systems.
The topology must support the intended application and its workload. Choosing a location is one part of that decision; capacity, access and maintenance also matter.
On-premises or customer-controlled hosting
An on-premises arrangement uses the organisation's own infrastructure. A hosted arrangement can use rented infrastructure under the agreed tenancy, access and network controls. In either case, identify the actual service locations and responsible operators.
The local-inference posture remains the same: the relevant inference service is within the deployment boundary. Hosting the environment on rented infrastructure does not imply sending tenant context to an external public model service.
Map the components
| Component | What to establish |
|---|---|
| Application and server services | Supported versions, endpoints, access and operating ownership |
| Database | Organisation context, credentials, backup, recovery and capacity |
| File/object storage | Supported storage arrangement, access, retention and restore procedure |
| Inference and retrieval | Model/service capacity, permitted context, endpoints and network path |
| Integration | External systems, exchange direction, credentials and failure handling |
| Observability and support | Available health information, logs, escalation and authorised support access |
An S3-compatible storage arrangement is part of the architecture. The selected provider and configuration must be supported by the implementation; compatibility should be demonstrated rather than inferred from a label.
Separate inference from business exchange
A process may read an external system or make a separately authorised outbound delivery. Those connections must be listed even when all AI inference remains local. Record what leaves, why, under which authority and what confirms receipt.
Follow representative requests through the configured endpoints. Agree any data handling, logging, retention and support arrangements. A location diagram alone does not establish that the full operation meets the organisation's requirements.
Plan continuing operation
Assign updates, version compatibility, configuration maintenance, secrets, storage growth and recovery responsibilities. Review backup coverage for both records and files, and demonstrate recovery under the agreed conditions.
Choose capacity for the expected people, data, reports, processes and AI use. A hosted and an on-premises environment may need different operating arrangements even when they run the same application capabilities.
Hand over one usable application
Verify a representative workflow, including a human decision, invalid input, denied action and unavailable dependency. Record the demonstrated environment, the results and any work still required for acceptance.
The integration-scoping guide develops source and exchange decisions. The catalogue-building guide develops the first P&C definitions. Architecture & Security gathers the wider evaluation questions.