Real Time Avator Voice Agent

This is a demo build pairing a real-time voice agent with a live AI avatar, answering spoken questions directly from a 200+ page official Microsoft annual report instead of returning a wall of text to read. It's a companion piece to this site's text-based RAG demo on the same report, built to test the same retrieval accuracy under a conversational, face-to-face interface instead of a chat window.

Under the avatar, the agent runs a retrieval pipeline against the actual report, so it answers from the document's real content grounded in specific pages rather than paraphrasing from general knowledge, while a real-time speech layer handles conversational turn-taking and a synced avatar renders the response as it's spoken. The engineering weight is in keeping retrieval, speech, and avatar rendering in sync so the answer arrives without an awkward lag between voice and face.

The result is a report Q&A experience that feels like a conversation instead of a search box: ask a question about the report out loud, and get a spoken, source-grounded answer back with a live face attached to it.