• 

  Clinical Copilot

Inspiration

Healthcare professionals spend too much time navigating fragmented systems instead of focusing on patients. Clinical data lives across EHRs, lab systems, insurance portals, and medication records, making it difficult to get a complete picture quickly. We wanted to build an AI-powered clinical assistant that acts like a “copilot” for healthcare workflows — helping clinicians retrieve insights, interpret patient information, and coordinate next steps in real time. The rise of interoperable healthcare standards like FHIR and agentic AI systems inspired us to combine intelligent orchestration with real clinical data access.

What it does

Clinical Copilot is an AI-powered healthcare orchestration platform built using multi-agent architecture and FHIR interoperability. It can: Retrieve patient demographics from FHIR records Summarize symptoms and assess urgency Suggest possible diagnoses Analyze medications for interactions and contraindications Interpret labs and observations Validate insurance and claims coverage Generate care plans and follow-up recommendations The system uses specialized AI agents working together under a single orchestrator to provide clinicians with contextual, explainable assistance.

How we built it

We built Clinical Copilot using:

Google ADK (Agent Development Kit) A2A agent communication SMART-on-FHIR integration Gemini models Python + Uvicorn Modular multi-agent architecture

The architecture includes: Root Orchestrator Agent Triage Agent Diagnosis Agent Medication Agent Lab Analysis Agent Insurance Agent Care Plan Agent Safety Validation Layer

FHIR scopes were configured for: Patient MedicationRequest Condition Observation Each specialist agent uses targeted tools and instructions while the orchestrator combines outputs into a unified clinical response.

Challenges we ran into

Some major challenges included: Designing safe healthcare AI workflows Managing agent orchestration and delegation Handling FHIR authentication and scopes Structuring modular agent communication Reducing hallucinations in medical reasoning Integrating multiple agents into a single A2A app Debugging ADK AgentTool and import issues Balancing performance with context size

We also had to carefully think about clinical safety and ensure that the system avoids giving overconfident medical recommendations.

Accomplishments that we're proud of

We are proud that we successfully:

Built a working multi-agent healthcare system Integrated real FHIR-compatible workflows Created modular specialist agents Implemented healthcare-oriented orchestration Designed scalable architecture for future production use Added medication safety and care planning capabilities Created an extensible framework for future healthcare agents Most importantly, we demonstrated how AI agents can collaborate safely and meaningfully in clinical environments.

What we learned

During this project we learned: How agent orchestration differs from traditional chatbot design The importance of interoperability standards like FHIR How modular AI systems improve reliability and scalability Practical challenges of healthcare AI safety How to structure multi-agent systems with ADK The importance of tool isolation and scoped permissions Strategies for reducing hallucinations in sensitive domains We also learned that healthcare workflows require explainability, safety checks, and human-centered design from day one.

What's next for Clinical Copilot

Next, we plan to: Add voice-enabled clinical interactions Support live EHR integrations Implement Retrieval-Augmented Generation (RAG) Add longitudinal patient history analysis Introduce clinician feedback loops Build multilingual support Add real-time emergency escalation detection Integrate imaging and radiology analysis Deploy secure cloud-native infrastructure Expand insurance and prior authorization automation

Our long-term vision is to create a trusted AI copilot that helps healthcare teams make faster, safer, and more informed decisions.

Built With

• gcp
• gemini
• google-adk
• python