How do you integrate AI capabilities into mobile applications?

Mobile AI integration balances on-device and cloud-based inference. On-device: use CoreML (iOS) or TensorFlow Lite (Android/cross-platform) for models that must run offline or require minimal latency. Cloud-based: call AI APIs through your backend (never expose API keys in mobile apps) for larger models and complex tasks. Architecture: mobile app calls your backend API, backend routes to AI services, processes results, and returns to mobile. On-device models: image classification (MobileNet, 4MB), text classification (DistilBERT quantized, 60MB), speech recognition (Whisper tiny, 39MB), and face detection (MediaPipe, 5MB). Cloud-based features: text generation (GPT/Claude via backend proxy), image generation (DALL-E/Stable Diffusion), and complex NLU. Implementation: implement a feature flag system to toggle between on-device and cloud processing, cache AI responses on device for offline access, implement background processing for non-urgent AI tasks, and manage model updates through app updates or dynamic model downloads. Performance: on-device models should initialize at app launch, use background threads for inference, and provide immediate UI feedback. Network optimization: batch API calls, compress request payloads, implement request queuing for poor connectivity, and design offline-first flows that sync when connected.