A voice-controlled clock that runs its command routing on-device: mic → Voxtral STT → a locally fine-tuned Laya router, cloud LLM only as fallback. Speaks French and English, including mixed.
The interesting part wasn't the app, it was the measurement. Zero-shot routing scored 91% in English but 62% in French, so I generated 2,500 labeled commands, fine-tuned Laya's multilingual checkpoint with RLCD on a Colab T4, and took French to 93%, then spent two more rounds chasing the confident mistakes that remained, and learned when to stop training and fix safety in plain code instead.
A voice-controlled clock that runs its command routing on-device: mic in, Voxtral STT, a locally fine-tuned Laya router, cloud LLM only as fallback. It speaks French and English, including mixed sentences.
The interesting part was not the app, it was the numbers. Zero-shot routing scored 91% in English but 62% in French. That gap was the whole project: a bilingual device that quietly works better in one language is not bilingual.
I generated 2,500 labeled commands and fine-tuned Laya's multilingual checkpoint with RLCD on a Colab T4. French routing climbed to 93%, then two more rounds chased the confident mistakes that remained.
A model that answers confidently in the wrong language is worse than one that says nothing.
The last round taught me when to stop training and fix safety in plain code instead. Some failure modes are not data problems; they are design problems wearing a data costume.