AI-Enabled Virtual Care with Digital Avatar Assistants
Date:
I delivered a talk at Amazon’s Image and Video Generation Workshop 2025 in Seattle, presenting our work at Amazon Health on building AI-enabled virtual care experiences using LLM-powered digital avatar assistants.
The talk covered the design and architecture of real-time, multimodal video avatar systems capable of conducting natural, task-oriented healthcare interactions. I discussed how large language models, low-latency speech pipelines, and expressive video avatars can be orchestrated into coherent end-to-end experiences, with a focus on healthcare pre-visit intake and digital provider interactions as a concrete application area. I also presented our evaluation of multiple approaches to avatar representation and synthesis, including video-based generation pipelines, diffusion-based models for high-fidelity synthesis, and 3D Gaussian splatting–based representations. The talk highlighted architectural trade-offs around latency, realism, turn-taking, and deployment constraints.
This work is described in more detail in the Video Avatars & Multimodal Experiences portfolio entry.
