Skip to main content

Common Workflows

Workflow 1 — Prototype to Production​

Use this workflow when you are building something new and need to validate before committing to wider deployment.

Collect samples (20–50 per class)
  ↓
Submit for training (minutes)
  ↓
Hot-swap to test device
  ↓
Validate in real conditions
  ↓
Retrain if needed (repeat above)
  ↓
Package as Personal SDK or Enterprise Batch
  ↓
Deploy to production devices

Key principle: Hot-swap is for learning and validation. SDK packaging is for distribution. Do not skip the field validation step — lab accuracy and real-world accuracy are often different.

Workflow 2 — Fleet Rollout​

Use this workflow when deploying an already-validated model to multiple devices or sites.

Start with validated model (from Workflow 1 or prior project)
  ↓
Create Enterprise Batch SDK
  ↓
Issue authorization keys for each site or customer
  ↓
Run batch-flash operation across device fleet
  ↓
Confirm activation status in model management dashboard
  ↓
Configure Agent Workflow per-site (if behavior differs by site)

Key principle: Authorization keys are issued before deployment. Confirm all keys are active before treating a rollout as complete.

Workflow 3 — Model Update​

Use this workflow when conditions change and the existing model needs retraining.

Identify the failure mode (what is the model missing or misclassifying?)
  ↓
Collect new samples targeting the failure mode
  ↓
Add samples to existing project (do not start a new project)
  ↓
Retrain (builds on existing training history)
  ↓
Validate against both original test set and new samples
  ↓
Deploy update using same path as original (hot-swap / batch)

Key principle: Retraining adds capability — it does not erase prior learning. If accuracy on original categories drops after retraining, the new samples may be unrepresentative. Investigate before deploying.

Workflow 4 — Offline Field Deployment​

Use this workflow for deployments where devices will operate without network connectivity.

Complete model training and validation with connectivity
  ↓
Package for target deployment path
  ↓
Configure Agent Workflow for offline operation:
  - Set detections to log locally
  - Configure queued alerts for next connectivity window
  ↓
Deploy to devices during maintenance window (USB or local network)
  ↓
Confirm inference is running on disconnected device
  ↓
On next connectivity: confirm logged events are synced

Key principle: Edge inference runs without connectivity. Only Agent Workflow actions that require external services (sending alerts, API calls) need connectivity. Design your workflow to buffer those actions if connectivity is intermittent.

Troubleshooting Common Issues​

Model accuracy is lower in the field than in testing
Likely cause: training images were collected under different conditions than real-world operation. Collect new images directly from the deployed camera in actual conditions and retrain.

Hot-swap deployment is not activating
Confirm the device is on a compatible firmware version. Check the device connection and try the deployment process again.

Agent Workflow alert not firing
Confirm the detection confidence threshold in your workflow configuration matches what the model is actually producing. Use the live test view to see real confidence scores before adjusting the threshold.

SDK key shows as inactive after issuing
The device must connect to activate a newly issued key. Confirm the device has had at least one network-connected session since the key was issued.