We built a realtime system for responsive voice AI in six months