Building a believable training conversation

I am contributing to a broader $5.2M Department of Defense-funded virtual patient grant project at Arizona State University. My work focuses on AI-driven virtual patients that present insomnia-related cases for therapist training and clinical education.

The goal is to create a conversational experience that behaves consistently like a patient within a defined scenario. The system must retrieve the right case context, preserve patient state across turns, and respond in a way that remains useful for a structured training session.

My role

I redesign and evaluate the conversational workflows that shape how the virtual patient interprets a therapist's question and produces a grounded response. The work combines RAG, semantic retrieval, prompt engineering, and agentic AI orchestration.

• Ground responses in relevant patient and scenario context.
• Improve semantic retrieval across varied therapist prompts.
• Coordinate agent responsibilities across the dialogue flow.
• Refine prompts for consistency, realism, and response quality.
• Evaluate failures and feed the findings back into the workflow.

How the workflow comes together

1. Interpret the therapist's input and current dialogue state.
2. Retrieve semantically relevant patient and scenario context.
3. Route the context through the appropriate agent workflow.
4. Generate a response constrained by the virtual patient profile.
5. Evaluate the response for grounding, consistency, and accuracy.
6. Use evaluation findings to improve retrieval and prompting.

Why it matters

Virtual patients can provide repeatable, accessible practice scenarios for therapist training. A well-grounded system gives trainees room to explore different conversational approaches while keeping the simulated patient's history and behavior coherent throughout the session.