Very cool! starred and on my reading list. Would love to chat and share notes, if you'd like

Also consider using Cerebras' inference APIs. They released a voice demo a while back and the latency of their model inference is insane.