One visit. Continuous care.
Google Cloud-native multi-agent care coordinator for chronic disease patients.
CareFlow turns one post-visit patient concern into an auditable care workflow:
Google ADK routes, Gemini reasons, MCP tools act, and AlloyDB AI remembers.
Safety gates supervise, while patients see what happened through a Care Receipt.
Sources: WHO NCD fact sheet, WHO hypertension report news, IDF Diabetes Atlas 2025.
CareFlow is a post-visit care coordinator for chronic disease patients across APAC personas. In the demo, Rajesh Sharma, a 63-year-old patient in Bangalore with type 2 diabetes and hypertension, asks a care question after a clinic visit. CareFlow routes that concern through the right agents, retrieves structured care memory, grounds external actions through tools, and returns a patient-readable result with proof of work.
CareFlow separates reasoning from action: ADK routes the patient concern, specialist agents reason over care context, and MCP/tool boundaries execute Gmail, Calendar, Maps, AlloyDB, notes, and notification workflows with visible proof.
Cloud Run hosts the API-based system. Google ADK and Gemini coordinate agents. AlloyDB AI, MCP, safety, observability, and healthcare data services provide the production substrate.
This view shows the broader build surface: frontend, backend, ADK runtime, Google Cloud services, deployment, observability, safety, and evidence.
The deployed service is a full-stack product, not a notebook demo: React/Vite for the patient experience, FastAPI for the backend, ADK for agent runtime, and MCP surfaces for scoped tool execution.
Dense full-stack appendix for technical reviewers
This appendix view is intentionally dense. It is useful for code review after the main deck/demo has already established the product story. It also covers the optional A2A discovery surface and outbound MCP/tool boundaries without adding another main-page diagram.
The UI makes agent work inspectable: dashboard context, multi-agent reasoning, grounded medical citations, human-in-the-loop confirmation, scheduling, doctor context, and care receipts.
- Open the Cloud Run app.
- Start with Rajesh Sharma, the default Bangalore persona with type 2 diabetes and hypertension.
- Ask:
My BP is high and I have a headache. What should I do? - Watch the system route through
root_agent, symptom triage, safety, data lookup, and caregiver escalation. - Check the Activity Feed and Care Receipt for concrete proof: what was checked, saved, sent, or scheduled.
- Ask:
Book my next cardiology follow-up at Apollo Hospital next Friday at 10am. - Verify that scheduling and location grounding are handled through the tool plane, not model memory.
- Switch persona/language to confirm APAC readiness and accessibility controls.
CareFlow is a hackathon prototype, not a medical device. It is designed to support care coordination, not replace clinicians.
This section is intentionally compact. It gives technical judges direct proof paths after they watch the demo and open the deployed app.
Evidence links: judge guide · architecture · observability · QA · GCP healthcare audit · fairness · i18n · Vertex Eval · MCP verification
CareFlow is a high-fidelity hackathon prototype. It demonstrates the production shape of a healthcare multi-agent system, but real patient deployment would still require clinical validation, privacy/legal review, durable HITL state, EHR partner integration, monitoring policies, and a formal medical safety process.
Built by Team Inference Engines
Google Cloud Gen AI Academy APAC Edition · Cohort 1