On-device keyboard prediction

Design an on-device + cloud ML inference system for a mobile keyboard's next-word/auto-correct prediction used by 500M devices, where predictions must run locally in under 20ms with no network dependency, models must update over the air, and personalization must respect on-device privacy (typed text shouldn't leave the device). Cover the on-device model, the update mechanism, how personalization works without sending raw data, and the cloud's role.