Skip to main content

Quick Start

This guide walks you through the complete process: training a custom model, validating it, and deploying it to hardware. No prior AI experience required.

What You Need​

  • A Hubosmart account
  • A supported edge device (see hardware overview)
  • 20–50 sample images per category you want to detect
  • A USB cable or local network connection to your device

Step 1 — Create a Project​

Log in and create a new project from the dashboard. Give it a name that describes what you are detecting (e.g., "Shelf Compliance — Store A" or "Line 3 Defect Detector").

A project holds your training data, trained models, and deployment history in one place.

Step 2 — Collect Training Data​

Navigate to the Data tab in your project. Add image categories — each category is one class the model will learn to recognize.

Tips for good training data:

  • Capture images in the actual conditions where the model will run: same lighting, same camera angle, same distance
  • Include variation within each class — different examples of the same condition
  • Aim for at least 20–30 images per class before training; more is better
  • Make sure negative examples (the "none of the above" class) reflect what the camera actually sees when nothing is happening

You can upload images from your computer or capture them directly through the interface if your device has a connected camera.

Step 3 — Train the Model​

Go to the Train tab and click Start Training. The training job runs on Hubosmart's cloud infrastructure — you will see accuracy metrics update in the browser as training progresses. This typically takes 2–5 minutes.

When training completes, review the accuracy results:

  • Overall accuracy tells you how often the model is correct across all classes
  • Per-class accuracy shows where the model is confident and where it struggles

If accuracy on a particular class is low, add more training images for that class and retrain.

Step 4 — Validate​

Before deploying, use the Test tab to check the model against images it has not seen before. Upload new images or use a live camera feed and observe the confidence scores.

A model ready for deployment should:

  • Correctly classify the expected condition with high confidence (>85%)
  • Not misclassify negative examples as positive detections

If the model is making systematic errors, add training data that corrects the specific failure mode, then retrain.

Step 5 — Deploy​

Select Deploy and choose your deployment path:

  • Hot-Swap — for testing on a single device right now
  • Personal SDK — for distributing to individual customers
  • Enterprise Batch — for deploying to a fleet of devices

For hot-swap, connect your device and follow the on-screen steps. The model will be active within a few minutes.

Step 6 — Configure Agent Workflow (Optional)​

If you want the device to respond automatically to detections — sending alerts, calling APIs, logging events — configure the Agent Workflow for your project.

Define the conditions that should trigger a response, and the action to take. No code required.

Next Steps​