Site sources
Cameras and local operating context.
Edge processing
Inference close to the source.
Site / cluster services
Configured aggregation and control.
Central oversight
Agreed events, evidence and management.
Deployment topology, not federated training. Offline behaviour must be specified for the delivery.
A deployment topology
The product materials describe a central office, subsidiary clusters and site-level edge servers. This is a distributed deployment pattern, not a claim of federated model training. Camera counts shown in an example diagram are not platform limits.
Processing and oversight
Place source acquisition and inference according to connectivity, latency and data locality. Specify which metadata, event evidence and control information are exchanged with the central environment.
Versions and local settings
Manage shared model/scenario versions while retaining justified site-specific zones and parameters. Record which version and configuration is assigned to each source.
Interrupted connectivity
Define local behaviour, buffering, retries and reconciliation. Confirm which services remain available during a disconnect and which actions require the central environment. The topology alone does not establish these guarantees.
Rollout and recovery
Use staged deployment, per-site validation, monitoring and rollback. Assign owners to central services, local nodes, network connections and operator workflows.
Separate result flows from configuration flows
| Direction | Content | Agree before deployment |
|---|---|---|
| Site → cluster / center | JSON events, snapshots and selected video clips | Scope, volume, latency, queue and retention |
| Management → execution | Model versions, policies and scenarios | Compatibility, testing, permissions and distribution procedure |
| Between analytical levels | Aggregated events and local reports | Aggregation rules and reference-data mapping |
| Corporate data store / BI | Agreed results and analytical data | Format, owners, access and retention |
Loss of Internet connectivity is different from losing camera access. If the processing node cannot read its source, it cannot analyze new frames. Specify diagnostics and any later processing of available recordings.
Third-party models shown in an architecture remain subject to package, category and runtime compatibility. Their presence in a diagram does not establish support for every artifact.