Tiny, always-on and fragile: Bias propagation through design choices in on-device machine learning workflows