Built a real-time conversational voice agent using LiveKit Agents, streaming STT, VAD, LLMs, and TTS; implemented interruption handling, acoustic/textual state classification, and state-aware response strategies. • Reduced measured voice-response latency from 7–8 s to 3–3.5 s by optimizing turn detection and classifier initialization; built per-turn observability for latency attribution and asynchronous event ownership.