Core ML, Vision, Natural Language, and on-device intelligence
On-device inference can reduce network dependence and data exposure, but local execution is not automatically accurate, small, fair, fast, or energy efficient.
Design the capability
- Place a versioned model behind a purposeful protocol so product rules do not depend on generated model types.
- Specify input preprocessing, output labels, confidence calibration, unsupported input, cancellation, and fallback.
- Measure accuracy by meaningful segments plus latency, memory, package size, and energy on supported devices.
- Keep a human correction path and record model version with derived results. Never present confidence as certainty.
Validation boundary
No model asset, Core ML compilation, Vision pipeline, Natural Language request, accuracy dataset, or physical-device benchmark was run.
Series navigation
- Previous: Part 105: HealthKit, WorkoutKit, and health-data design
- Next: Part 107: StoreKit, subscriptions, offers, and entitlement state
- Series index: Zero to iOS Hero