Looks good. Can you provide some info on what the software stack looks like? Is the device programmable or is it inference only from a fixed selection of models?

You lost me at 35dB.