You don’t have to stand up infrastructure to start. If a teammate has already deployed AcuSight and provisioned devices, you can work entirely in Projects — annotate, version, train, evaluate, promote.
Your path
1
Learn the data model
A handful of nouns — project, batch, dataset version, and model — carry the whole workflow. Skim the Glossary first; it’s quick.
2
Organize your data
Incoming device images land in the Raw Pool as batches. Assign them to a project and move them through the pipeline. See Projects & batches.
3
Annotate
Label images in the browser — draw boxes, assign classes. See Annotation.
4
Version a dataset
Create a snapshot — an immutable dataset version with train/validation/test splits, plus optional preprocessing and augmentation steps. See Dataset versioning.
5
Train and evaluate
Train a model against a dataset version: pick an architecture (YOLOv5 / YOLOv8 / YOLO11) and size, then watch accuracy and loss metrics stream in live. See Training & experiments.
6
Deploy to devices
Deploy the model you’re happy with to a group of devices. See Model evaluation & deployment.
Good to know
- Datasets are immutable. Once you snapshot a version, its contents never change — new labels or batches mean a new version. That’s what makes a training run reproducible.
- One model per dataset version. Each snapshot trains one model; retraining means creating a new version. It keeps “which data produced this model” unambiguous.
- Training is object-detection-first. Jobs run a YOLO backend and export to a range of formats (PyTorch, ONNX, TFLite, and Synaptics NPU for edge hardware).
- Classes are per-project. You define the classes and tags a project’s models can detect — each with its own color — on the Classes & Tags page.
Start here
Build a defect detector
The end-to-end tutorial — data to deployed model — for your exact workflow.
Model lifecycle
How champion and baseline models fit together.