Add On-Device AI with Gemini Nano

Assess the app's task and device support before promising offline AI. Start with Google's official ML Kit GenAI documentation: https://developers.google.com/ml-kit/genai and the selected API's Android setup guide. Gemini Nano via AICore and downloadable MediaPipe models are distinct approaches; do not label an arbitrary local model as Gemini Nano.

Choose an integration

Inspect the Expo SDK, native build and supported Android versions. Select an official ML Kit task API or Prompt API where currently supported, then plan a narrow Kotlin bridge through Expo Modules. Verify documented dependency versions, permissions, model availability and licensing before proposing changes. Third-party wrappers require a separate maintainer, source and compatibility review; do not automatically install an unverified package.

Check feature status using the selected API's documented runtime availability method. Handle unavailable, downloadable, downloading and available states. A phone model name alone does not prove API availability. Model downloads may require connectivity, storage and user approval; describe this before starting them.

Privacy and fallback

Keep prompts and outputs out of diagnostics by default. Test whether the chosen integration sends telemetry or data before describing it as private. On-device processing does not remove applicable AI content, disclosure or reporting requirements.

For unsupported devices, show a useful unavailable state. A cloud fallback is a separate user choice: identify the provider and data transferred, obtain applicable consent, authenticate through a backend and enforce rate limits. Never send a prompt to the cloud automatically after a local failure. Never place provider credentials in mobile code.

Deliver and verify

Produce an integration plan with exact dependencies, supported API states and privacy boundaries. Test supported and unsupported physical devices, denied downloads, low storage, cancellation, error handling and a network-disconnected inference after model readiness. Record measured latency and output quality without invented device coverage or speed estimates.