Can a small but highly specialized language model handle your app's instructions?

In this session, I'll show how I built exactly that into Speak2Act, an open-source Android app that turns voice into transactions. Whisper transcribes speech on the phone, and rules still handle the common phrasing in microseconds. Everything the rules can't parse goes to FunctionGemma 270M, fine-tuned for one job and running on-device with LiteRT-LM. The cloud becomes the last resort, not the default, and yes, the demo runs in airplane mode.

You'll see how to:
• choose a runtime and a model small enough to ship
• get structured output with tool calling, using one contract for the phone and the cloud
• keep a small model honest: the model proposes, grammar verifies, the user confirms
• fine-tune with two LLMs, teacher and student.
• convert, quantize, pin and ship a 284 MB model like any other release artifact
• measure what matters: accuracy, latency, memory and download size

You'll leave with a recipe you can reproduce, because the app, the Colab notebooks and the fine-tuned model are all public on GitHub and Hugging Face. And with one lesson that applies far beyond payments: small and specialized beats big and general.