AI Vtuber: A Multi-Agent LLM Debate Streaming System

Overview A single LLM answering a question sounds authoritative whether or not it is right — hallucinations are hard to notice precisely because nothing contradicts them. This project takes the opposite approach: instead of one assistant handing down an answer, several agents are given opposing positions and made to argue , live, while the audience watches and joins in. The friction between them is the point; it gives a viewer a handle for scrutinizing the claims themselves. The result is an AI Vtuber debate stream on YouTube: agents collect information from the live web, argue in character, and respond to viewer comments in real time. Backend LLM — Gemini API on Vertex AI, using gemini 2.0 flash thinking exp where extra reasoning helps. Agent orchestration — LangGraph, so the debate flow (research → statement → rebuttal → summary) is an explicit graph rather than one long prompt. Live information gathering — browser use drives a real Chrome instance, so an agent can go look something up mid debate instead of relying on its training data. Memory — past utterances are embedded into ChromaDB and retrieved (RAG) so agents stay consistent with what they already argued. Infrastructure — the GPU bound parts run on GCE; the UI runs locally. Splitting it this way keeps the interface responsive and the cloud bill bounded. Frontend & streaming Unity drives the avatars — expression and gesture animations are triggered from the agent output, so the debate reads as embodied rather than as a wall of subtitles. AivisSpeech for speech synthesis, with lip sync driven by the audio. WebSocke…