Case Study + Deep Dive: Telemedicine Support Agents with LangGraph/MCP - Dan Mason

Introduction and Overview 00:00

• Dan Mason expresses excitement about returning to the event after attending last year, sharing insights from his work on telemedicine support agents.
• He introduces the workshop's focus on agent workflows in healthcare, specifically highlighting telemedicine use cases.

Project Background 03:13

• The client, Aila Science, focuses on women's health, particularly in managing early pregnancy loss.
• Patients receive medication after a miscarriage and must manage their treatment at home, which can be challenging without direct support.
• Aila’s system uses a text messaging interface to help patients navigate their treatment.

System Architecture and Tools 06:30

• Stride, the custom software consultancy, built the telemedicine workflows using a combination of traditional software and AI tools, particularly Langraph and Langchain.
• The architecture accommodates patient data privacy concerns, adhering to HIPAA regulations.
• The team consists of software engineers, a designer, and Dan Mason, who manages the Langraph side of the project.

Agent Workflows and Evaluation 12:10

• The agent system aims to automate responses to patient inquiries while ensuring human oversight.
• A self-evaluation function was designed to flag complex situations for human review, enhancing the workflow's efficiency.
• The goal is to achieve a 10x increase in patient handling capacity while maintaining quality care.

Patient Interaction Examples 18:06

• The system captures patient interactions, maintaining state and context through anchors and scheduled messages.
• Dan illustrates how the system engages with patients, tracking medication schedules and responding to queries in a flexible manner.
• The agent can adjust based on the patient's responses, such as updating time zones or treatment progress.

Handling Complexity and Errors 36:28

• The system is designed to recognize and manage complex patient interactions, ensuring that patients are guided back to their treatment pathways.
• Low confidence responses from the AI trigger human review, allowing for quality control in patient interactions.

Scalability and Future Development 44:08

• The system is built to support multiple treatments without extensive coding, allowing for rapid adaptation to new patient needs.
• Dan discusses the potential for scaling the software to handle larger patient populations effectively.

Final Thoughts and Q&A 70:00

• Dan reflects on the development process, acknowledging the challenges faced with integrating new technologies and ensuring reliability.
• The workshop concludes with a Q&A segment where Dan addresses various questions about the system's architecture, decision-making processes, and evaluation metrics.

Conclusion 116:00

• Dan thanks the audience for their participation, expressing hope that his insights will inspire improvements in telemedicine workflows.