CareBand — 5-Agent Healthcare Coordination System
CareBand is a 5-agent healthcare coordination system built for Track 3: Regulated & High-Stakes Workflows. It addresses a critical gap in chronic patient care: 1.4 billion patients worldwide, 50% miss medications regularly, and families find out 4–6 hours too late when emergencies happen. The system deploys five specialized AI agents that collaborate exclusively through a central message bus. Triage detects symptoms and assigns risk levels. Medication checks drug interactions and missed-dose severity. Guardian makes escalation decisions (MONITOR, ESCALATE_SCHEDULED, or ESCALATE_IMMEDIATE). Family alerts caregivers in real time via SMS, WhatsApp, or in-app notifications. Compliance logs every decision with immutable audit trails for GDPR and HIPAA. Every agent communicates through structured JSON handoffs — not chat, but real task delegation with context sharing. Band serves as the core collaboration layer: Triage sends to Band, Medication reads from Band, Guardian reads from Band, Family reads from Band, and Compliance reads all messages from Band. Remove Band, and the system collapses. The architecture is fully implemented with real-time SSE streaming, fire-and-forget message routing, and full pipeline execution. The current prototype uses a local message store that mirrors Band's exact routing logic, with live Band API integration pending OAuth token configuration — a transport-layer switch requiring zero code changes.
Category tags:
"CareBand presents a compelling 5-agent healthcare coordination system targeting a high-stakes, real-world problem: medication adherence and emergency escalation for chronic care patients. The domain choice is strong and the agent architecture is well-reasoned, with clear division of responsibility across Triage, Medication, Guardian, Family, and Compliance agents. The Band message bus concept is implemented thoughtfully — structured JSON handoffs between agents, real-time SSE streaming, and a centralized routing table that maps each agent's output to the appropriate downstream recipient. The demo video effectively showcases a patient scenario (Mary Smith, 103F fever, missed insulin) flowing through the full workflow with a clear visual of routing steps and escalation decisions. The compliance layer (GDPR/HIPAA audit logging via an immutable decision trail) adds meaningful enterprise-readiness to the concept and demonstrates awareness of regulatory constraints in healthcare AI. The multi-language support via the LANG field is a practical touch for wider deployability. However, the Band SDK integration is noted as incomplete in the README — the system currently uses a local message store as a substitute for the production Band API due to an OAuth authentication issue. While the architecture is designed to switch seamlessly, the live Band integration was not demonstrated, which limits the score on technology application for this hackathon specifically. The codebase is primarily TypeScript (96.6%) and deployed on Replit, which is a reasonable rapid-development choice. The agent memory directory and structured workspace layout show organized project thinking. Presentation quality is high: the animated orbital visualization of agents in the demo is visually impressive and communicates the multi-agent coordination concept clearly. The README is well-structured with the routing table and agent flow documented explicitly. Key improvements for future iterations: completing the live Band OAuth integration, adding persistent patient profiles, and building out real SMS/WhatsApp notification delivery would elevate this from a prototype to a production-ready system."
Dharma Singh
Senior Development Manager
For Innovators & Creators