Revamp architecture for video chatbot

What we've maxed out in current build: * Voice quality – OpenAI Realtime (best in class) * Reliability – retries, timeouts, backpressure all solid * KB search gap – filled with natural filler speech * Language – English + Hindi auto-detect working Two hard ceilings we've hit (platform limits, not bugs): * Lip-sync lag: 200–400ms — this is HeyGen's architectural ceiling, can't be fixed without switching approach * Voice accent: sounds American — OpenAI Realtime doesn't offer Indian accent natively Bottom line: Current build is a solid . We've squeezed everything possible out of this architecture. To go the best version need rework work: * Switch to text-based avatar speech for more better lip-sync * Plug in ElevenLabs/PlayHT Indian voice clone * Evaluate replacing HeyGen with Tavus or D-ID — they have native Indian avatars + built-in TTS, tighter sync, one system instead of three since we have other priorities too LMK Additional requirements: 1.⁠ ⁠Can we use an Indian avatar or p