Sonny runs on a lean, real-time voice stack: browser mic capture, VAD to know exactly when the user has stopped speaking, Groq's whisper-large-v3-turbo for transcription, Groq's llama-3.3-70b-versatile for reasoning, and Cartesia Sonic for low-latency speech output. The whole experience is wrapped in a web-native orb and a minimal conversation loop so it feels alive, immediate, and soulful. Built by Deepanshu.