Manufacturing: 750 component classes, ~1000 labeled examples per class. Build a photo-based component recognition app as an MVP within a few working days. Which approach do you take?

Cloud Vision AutoML is the most suitable choice for this scenario because it allows for rapid development of custom image classification models with high accuracy, even with a relatively large number of classes and examples. Its automated machine learning capabilities significantly reduce the time and expertise required, making it ideal for an MVP within a few working days. Reducing the dataset would likely degrade model performance and is unnecessary given AutoML's capabilities. Cloud Vision API is a pre-trained general-purpose API and does not allow for custom model training with specific component classes. Training your own model with transfer learning would offer more control but would take significantly longer than "a few working days" to implement, fine-tune, and deploy, especially for 750 classes.