Real-Time Avatar Launch Checklist

One-page engineering reference: the five-stage conversational loop (ASR->LLM->TTS->avatar synth->WebRTC); the ~200 ms turn-taking target and the ~900 ms time-to-first-frame budget; the join-the-room architecture (send audio only, avatar publishes video, RPC for interruptions); the 2026 buy-vs-build map (Tavus, HeyGen, Simli, bitHuman vs MuseTalk + LivePortrait + NVIDIA ACE); the EU AI Act Article 50 disclosure deadline (2 August 2026); and a 7-item pre-launch checklist.