Puffin

Puffin is a text-to-FHIR model with an intelligent orchestration agent for Electronic Medical Records (EMR) and Hospital Information Systems (HIS). Clinicians type plain language requests, and the system converts them into FHIR (Fast Healthcare Interoperability Resources) or HL7 messages that connect directly to the EMR. The agent can also trigger specialized tools for advanced workflows.

Why this is needed
Current EMR systems are:
• Clunky: Slow and unintuitive interfaces
• Isolated: Limited connection to external systems such as payers or evidence databases
• Data-rich, insight-poor: Large datasets with little real-time decision support
• Static: Rigid automations that cannot adapt to new workflows

How it works
• Puffin: Text interface for natural language commands
• Puffin Agent: Orchestration layer that interprets intent, selects tools, and routes requests
• Current tools: Discharge summary maker, financial estimation, insurance pre-authorization via browser or email, evidence lookup, and scribe assistant
• Puffin FHIR: Core text-to-FHIR model that produces structured, standards-compliant FHIR or HL7 messages
• FHIR/HL7 Interface: Connects to EMR/HIS and terminology servers for code normalization such as SNOMED, LOINC, and ICD-10
• Polling Worker: Monitors EMR/HIS changes and triggers follow-up actions

Potential expansion with MCP
Model Context Protocol (MCP) enables secure, dynamic connection to new tools without custom integration. Potential additions include pharmacy ordering, imaging PACS viewers, remote patient monitoring dashboards, clinical trial matching, specialist scheduling, and lab ordering systems.

Key innovations
• Text-to-FHIR intelligence: Removes the need for manual EMR navigation and form completion
• FHIR-first design: Compatible with most modern EMRs
• Agentic orchestration with MCP scalability: Expands capabilities on demand while keeping workflows seamless and compliant

Our work started on Saturday right after problem statement release. We spent several hours planning and initially failed to build a full on MCP text-to-FHIR with additional tooling. It failed and didn’t deploy well. It was only at 6pm that we decided to scale down and only implement a direct REST API Agentic workflow with text-to-FHIR integration. We completed our work at 4am.

Disclaimer: We are using an open EMR system with FHIR server by Canvas Medical. One of our team member is a doctor has a lot of experience with EMR/HIS data and healthcare interoperability from his current startup and professional work. Hence the background knowledge for interoperability and clinical knowledge was there from the beginning.