Investigating Video Impression of Real Time Conversational Agent for Remote Sessions Using LLM

Large Language Models (LLMs) have rapidly advanced in their ability to simulate intelligent, adaptive agents across a wide range of domains. While much of the existing work focuses on task-specific chatbots or standalone AI assistants, the integration of LLM-driven agents into real-time video conferencing remains largely unexplored. In this project, we present AIVA (AI-based Virtual Avatars) a multi-client video conferencing system that enables users to initiate live sessions with AI agents capable of adopting dynamic roles such as teacher, coach, or mentor based on user-provided descriptions. The system architecture combines LLM-powered reasoning for conversational intelligence with 3D avatar rendering to deliver a visually embodied, interactive agent experience. WebRTC is used to maintain uninterrupted real-time communication between clients and AI agents. When a user initiates a session with a prompt like “Act as personal fitness coach,” the system employs targeted prompt engineering to shape the agent's tone, behavior, and domain knowledge to match the requested role. The platform also includes essential features such as session control, avatar behavior customization, and conversation history management. Preliminary evaluations across academic tutoring, emotional support, interview preparation, and latency testing indicate that AIVA can produce highly engaging, human-like interactions in a live multimedia environment. These results high- light the system's strong potential for scalable, intelligent, and personalized virtual communication technologies, particularly in educational, advisory, and training-oriented applications.