The rise of Large Language Models (LLMs) has transformed human-computer interaction, developing conversations with machine seems more natural and transparent. Embodied Conversational Agents (ECAs), or 3D avatars, extend this work by providing AI a visual presence. This paper presents a complete, end-to-end architecture for a locally hosted, real-time voice assistant powered by a 3D avatar. The system integrates real-time Speech-to-Text (STT), Text-to-Speech (TTS), and a web-based 3D talking head developed on a modular backend with persistent client-server communication. The conversational intelligence is driven by a Retrieval-Augmented Generation (RAG) model connected to a Pinecone vector database, enabling contextual and domain-specific responses. The proposed design demonstrates a practical approach to creating efficient, low-latency conversational agents deployable on consumer hardware.