lukaskris/litert-lm-mobile-android ? reverse-engineered prompt

Build me an Android app that feels like a premium on device AI chat app.

I want it to use LiteRT LM with Gemma 4 as the main model, and let me chat with text, attach an image from the gallery, or record audio in the app and ask questions about it. It should stream answers smoothly in a modern dark blue and soft gray UI, and show live performance info like time to first token, tokens per second, and token count in the top area.

Make it smart about performance too. It should automatically try the best backend it can find on the phone, starting with NPU, then GPU, then CPU, and handle different Android chip makers without me needing to think about it. Add a simple way to switch models in settings, and make it work with models copied over by ADB from the computer instead of forcing an in app download.

If you need to look up current docs online while building it, do that.