customvision.ai (Azure AI Custom Vision)

Upload your own labeled images or use assisted tagging, train the model, then read precision and recall metrics on a held-out set before deciding to publish.

Handle two different problems: multiclass or multilabel image classification, and object detection that returns bounding boxes for recognized objects with confidence scores.

Published models expose an HTTP endpoint you can call from any language or platform; send an image and receive predicted tags, probabilities and boxes as JSON.

Export trained models for CoreML, TensorFlow, ONNX, Docker and other formats when you need to run inference on-device or in your own infrastructure.

Keep multiple iterations of a model side by side, compare their performance, and promote the best one to production without retraining from scratch.