Standardized Patient Avatar for Reflective Communication Practice (SPARC-P)

Clinicians regularly navigate complex, emotional conversations. However, opportunities to practice these skills are limited. Traditional training with live standardized patients is effective but resource-intensive, requiring significant time, cost, and coordination.

SPARC (Standardized Patient Avatar for Reflective Communication Practice) offers a new path forward: a virtual, AI-powered patient that enables clinicians to practice anytime, anywhere, with instant objective feedback that supports ongoing skill development.

Project Timeline

Planning for SPARC started in July 2025 and work on the project will proceed in the six stages outlined below.

Stage 1: Planning & Onboarding (Complete)

Internal kickoff meeting
Establish project charter, timeline, and workflows
Define technical requirements and learning goals
Begin character concept development
Outline web and data infrastructure needs

Stage 2: Initial Design & Prototyping (In Progress)

Develop character profiles and dialogue tone guides
Create wireframes for web interface
Map learner experience and design assessments
Begin early AI agent scripting and risk assessment

Stage 3: Alpha Build Development

Integrate character designs into agent framework
Build and deploy alpha version of training platform
Implement basic conversational AI agents
Define data capture structure and feedback loops

Stage 4: Internal Testing & Risk Review

Conduct functionality and interaction testing
Perform AI risk assessments
Analyze alignment of agent dialogue with instructional intent
Gather internal team feedback

Stage 5: Refinement & User Simulation

Refine agents for contextual depth and realism
Conduct learner simulations and usability testing
Adjust instructional flow and platform design
Monitor system performance and data output

Stage 6: Finalization & Reporting

Final QA and bug resolution
Prepare documentation and final deliverables
April 2026 Presentation
Submit report
Team debrief and planning for full PCORI integration/Scalability

Technical Details

As you might expect, creating a fully functional AI avatar requires careful orchestration of many systems. The following diagrams outline keys processes used the bring the avatar to life!

Diagram showing how a generic multi-agent system interacts with the user.

Multi-Agent AI

This diagram shows how a multi-agent system typically interacts with the user.

Diagram showing AI training document processing

Processing the Training Documents

Documents used to train the AI model may contain sensitive information and must be processed. All curated files are converted to text, and then they are then fed to a helper application that removes personally identifiable information.

Diagram showing how text documents are converted to JSON and then embedded into a vector database using LangChain and Chroma.

Create a Specialized Index

Our AI model uses a special multi-dimensional index called a vector database. This vector database is used to identify patterns within data set that will be useful when processing and generating responses.

In our case, we take the text documents processed in the previous step and convert it into JSON. From there, specialized programs called LangChain and Chroma will generate the vector database.

Digram showing how vector database is combined with an open-source LLM into a single knowledgebase which is then used to train each agent.

Combine Vector Database with an Open Source LLM

Large Language Models such Open AI’s GPT or Meta’s LLaMA have been designed to create natural sounding text output. When combined with the specialized vector database, the result is a specialized AI model able to respond from a particular point of view.

Creating the Avatar

This diagram shows how the parent avatar is created. The process uses mesh tools like Blender which are fed to Character Creator 4 to create a visual representation of the character.

Voiceover audio is fed into NVIDIA’s Audio2Face tool to generate the animation required for facial motions when speaking. The character and animation then get fed into iClone 8 to refine the animation. The Animated Character can then be exported for use on the frontend.

User Interactions With the AI Avatar

With the data now in place, the AI avatar is ready to interact with users. This diagram shows how the user input from their microphone is translated into a text format which can then interact with the rest of the AI model. In the end, the avatar’s audio and visuals are sent back to the user.

Diagram showing what happens on the backend hosted on HiPerGator. When the user speaks, his words are transcribed using NVIDIA Riva. Next, they are filtered to remove inappropriate or dangerous content and sent to the supervisor agent.

HiPer Gator Processes Incoming and Outgoing Data

As new responses come in from the user, the HiPer Gator supercomputer goes to work! New input is filtered for safety and fed to the supervisor agent. The supervisor agent passes this data on to the parent and coach models.

The parent model generates a response, and then it feeds it back to the supervisor agent. The supervisor then feeds that text into the the animation engine that calculates the animation needed for the text that be said by the avatar. This info gets passed back to the frontend.

The first chart shows an overview of the entire process for generating the interactive character. The second chart shows the specific implementation of backend on housed on HiPer Gator.

Diagram showing how the user's voice is converted to text that can be processed and how the AI uses the service to produce audio responses.

Processing AI Avatar Audio

The AI Avatar will depend on RIVA AI Services to handle audio processing. As the user speaks the microphone signal is converted to text by automatic speech recognition (ASR). This text gets back to HiPer Gator to be processed by the Supervisor Agent. Think of this as how the AI agent “listens.”

RIVA also handles converting messages from the supervisor agent into audible speech. This process is essentially the reverse of the “listening” process. RIVA receives text data from the supervisor agent about what to say. This text gets run through a text to speech (TTS) engine to produce the audio output. This is sent to the supervisor agent to sync the audio with the animation before being delivered to the user.

Poster presented at AI Days for SPARC Project

Download Poster

Meet the Team

Carma Bylund, Ph.D.

Co-Principal Investigator

Dr. Bylund is a Professor and Associate Chair of Education in the Department of Health Outcomes & Biomedical Informatics (HOBI) and the Assistant Director of the Cancer Training and Education Program at the UF Health Cancer Center (UFHCC). As an implementation scientist, her research focuses on communication interventions for clinicians, patients, and caregivers.

Jason Arnold, Ed.D.

Co-Principal Investigator

As the Director of E-Learning, Technology, and Communications—and serving as the Senior Communicator for the College of Education—Dr. Arnold leads a diverse team comprised of instructional designers, online student services support, web designers, software and database programmers, videographers/editors, and graphic designers to support teaching and learning and advance the mission of the college and university.

Stephanie Staras, Ph.D.

Co-Principal Investigator

Stephanie A. S. Staras, M.S.P.H., Ph.D., is a professor and Associate Chair for Faculty Development in the Department of Health Outcomes and Biomedical Informatics. She is also the Associate Director of the Institute for Child Health Policy and co-lead for the UFHealth Cancer Center’s Cancer Control and Population Science Program.