We do some preprocessing on-device and run our more complex models in the cloud. It’s similar to Alexa which always listens for the trigger word, but only streams to servers for the few seconds afterwords. While some image data is processed in the cloud, it’s never stored.