Skip to main content

Model Management

Training a model is one event. Managing it over time — tracking versions, controlling who has access, knowing what is deployed where — is an ongoing operational concern. Hubosmart's model management layer handles this without external tooling.

Training History

Every training run is recorded. You can compare accuracy metrics across versions, identify when a model improved or regressed, and return to a previous version if a retrain produces worse results.

When you retrain to address new conditions — a new product SKU, a new defect type, seasonal crop variation — the history of prior versions remains accessible. Deployment decisions are informed, not guesswork.

Version Control

Models are versioned at the point of export. When you deploy a model, the platform records which version was deployed to which device or SDK. If behavior in the field changes, you can identify exactly which model version is running and compare against the training record.

Version eventWhat is recorded
Training runDataset snapshot, accuracy metrics, training configuration
Model exportVersion tag, timestamp, destination (hot-swap / Personal SDK / Enterprise Batch)
SDK issuanceKey ID, issued to, credit cost, activation status
Key revocationTimestamp, associated device or customer

SDK Key Management

For Personal SDK and Enterprise Batch deployments, access to your trained model is controlled through the key management interface.

Credit-based issuance: Each SDK key issued costs credits from your account balance. This makes distribution cost predictable — you know what each deployment costs before you issue it.

Key lifecycle:

  • Issue a key when onboarding a customer or activating a device
  • Monitor activation status to confirm deployment succeeded
  • Revoke a key when a customer relationship ends or a device is decommissioned

Keys are bound to the SDK they authorize. Revoking a key disables that specific activation without affecting other customers or devices.

Fleet Visibility

For Enterprise Batch deployments, the platform provides an overview of the deployment fleet: which devices have received which model version, activation status, and which keys are currently active.

This removes the need to track fleet state in a spreadsheet. When you issue an update — a retrained model bundled into a new SDK — you can see which devices are on the old version and which have received the update.

Keeping Models Current

Hubosmart is designed for iterative model improvement. When your operating environment changes:

  1. Add new training samples reflecting the changed conditions
  2. Retrain — minutes for typical classification tasks
  3. Export and deploy to existing devices using the same path (hot-swap for fast iteration; batch for fleet updates)

The platform handles the mechanics. The decision about when to retrain and what new samples to add belongs to the people who understand the domain.