Digital Virtual Assistant Based on Expert Hybrid Medical Large Language Model

Digital virtual assistant based on Mixture-of- Experts (MOE) is an integrated product that incorporates artificial intelligence large models and digital avatar technologies. It collects user data on lifestyle, diet, exercise, and mental state via voice interaction and smart devices, providing real-time health assessments and adjusting recommendations based on daily activities and feedback. To achieve these functions, this paper proposes the following methods: (1) we developed a large language model with sparse MoLoRA (Mixture of Low-Rank Adaptation) for multi-task medical learning, enabling professional medical dialogues across diverse tasks without specific annotations, (2) a real-time consultation process for digital avatars integrates voice recognition, large model Q&A, text-to-speech, facial animation, and full-body rendering, offering realistic and interactive experiences, and (3) we optimized the digital human image to align with Asian aesthetics while providing customization options to meet diverse user preferences. Experiments show our solution outperforms competitors, providing an efficient, user-friendly virtual assistant with strong market potential, deserving wide promotion.