Inspiration
Every day across America, hospitals discharge unhoused patients with nowhere safe to go. We were shocked to learn that over 50-70% of homeless patients in U.S. hospitals are discharged to unsafe or unstable environments without a clear care plan.
The current process is slow, fragmented, and relies on phone tag between hospitals, shelters, transport providers, and social workers. Case coordinators spend 4-5 hours per case, making dozens of phone calls, often only to find shelters are full or transportation isn't available.
We built CareLink to solve this critical problem by creating an intelligent, voice-enabled coordination system that ensures every unhoused patient leaving a hospital has a verified shelter placement, coordinated transport, and connected social worker for continuity of care.
What it does
CareLink is a decentralized, multi-agent AI system that automates the entire discharge planning workflow for unhoused patients:
9 Specialized AI Agents (Powered by Fetch.ai)
- Coordinator Agent: Orchestrates the entire workflow https://agentverse.ai/agents/details/agent1qf5tefy3dd9jwtwl8s0ht0aq6yn4du7ntvwdhhcfxqwwjas7fwcpsme6g35/profile
- Parser Agent: Extracts structured data from discharge documents using LlamaParse + Gemini AI https://agentverse.ai/agents/details/agent1qthwalvx5c757u0gsdrx6xfu9lf0y9cc6jce8ddvx87cy2w9a7mgv0rthgr/profile
- Shelter Agent: Matches patients with available shelter beds using real-time data https://agentverse.ai/agents/details/agent1qd4kn4syvtlypa20aq8pct4d6wa6fke77su5tjqsj2ymgr2n0klew447j3s/profile
- Transport Agent: Schedules accessible transport (Paratransit, Lyft Access, Uber WAV)\ https://agentverse.ai/agents/details/agent1q0vm9tl5d5tjpj3d7qqk692jz0y9f486twxmjzfv8eng288gqf0pgjt83nd/profile
- Social Worker Agent: Assigns case workers and schedules follow-ups https://agentverse.ai/agents/details/agent1qfxx6ry60e9clra6vgyp76q0uvfrkc6n0kga4cln3f7c6fnl44wgujvphnd/profile
- Resource Agent: Coordinates food, hygiene kits, and clothing https://agentverse.ai/agents/details/agent1qfz242vsc6hr3ck48kmef73hlj39538ksgt2w4v3fpawyjp8q2yrwr2hdfp/profile
- Pharmacy Agent: Ensures medication continuity post-discharge https://agentverse.ai/agents/details/agent1qdaq0qhnsgt3f3p0sk0f5642fyceqfh44739prhv0x9x828ay0a6vh4dgen/profile
- Eligibility Agent: Verifies benefits (Medi-Cal, GA, SNAP) https://agentverse.ai/agents/details/agent1qf0phyg2dqm4q5nee9jx8ukzlry5k3x388rqzdp24kksujmc2nzysa398tf/profile
- Analytics Agent: Tracks system performance and outcomes https://agentverse.ai/agents/details/agent1qg0g6fm790qu6rcl6tqulk5gvvvq0ma5a4tzcwzkdzp3dr8tneqdjldgltj/profile
Voice-First Communication (Powered by Vapi)
- Automated shelter availability verification calls
- Social worker outreach and coordination
- Transport provider scheduling
- Post-discharge patient check-ins
- Accessible for low-tech organizations
Real-Time Intelligence (Powered by Bright Data)
- Live shelter capacity data from San Francisco HSH and community sources
- Up-to-date resource listings (food banks, medical respite, outreach centers)
- Transport provider schedules
- Continuously refreshed community resource database
Intelligent Document Processing
- AI-powered PDF parsing with LlamaParse
- Automatic form autofill from discharge documents
- Gemini AI for data extraction and structuring
- Confidence scoring with manual review flagging
How we built it
Frontend:
- Framework: Next.js 14 with TypeScript
- UI/UX: Tailwind CSS + Framer Motion for smooth animations
- Maps: Mapbox GL JS + react-map-gl for interactive shelter mapping
- Components: React 18 with modern hooks and state management
- Icons: Lucide React for consistent iconography
Backend:
- API Framework: FastAPI (Python)
- Agent Framework: Fetch.ai uAgents for decentralized coordination
- Database: Supabase (PostgreSQL) for persistent case storage
- Voice AI: Vapi for automated phone communications
- Document Processing: LlamaParse for PDF parsing
- AI Intelligence: Google Gemini 1.5 Pro for data extraction
- Web Intelligence: Bright Data for real-time shelter/resource data
DevOps & Integration:
- HTTP Client: httpx for async API calls
- Tunneling: ngrok for webhook handling during development
- Testing: pytest for backend testing
- Code Quality: black, ruff, mypy for Python; ESLint for TypeScript
Challenges we ran into
Agent Communication Complexity Challenge: Coordinating 9 independent Fetch.ai agents without message loss or race conditions
Solution: Implemented a robust coordinator pattern with:
- Message queuing and retry logic
- State synchronization via Supabase
- Unique message IDs to prevent duplicates
- Timeout handling for unresponsive agents
Timeline Visualization Challenge: Displaying complex multi-agent workflows in an intuitive timeline
Solution:
- Used Framer Motion for smooth animations
- Built hierarchical status tracking (agent → task → subtask)
- Implemented collapsible sections for detailed logs
Accomplishments that we're proud of
9 Production-Ready Fetch.ai Agents - Fully functional multi-agent system with real coordination
Vapi Voice Integration - Successfully automated shelter verification and social worker outreach calls
Bright Data Intelligence - Live web scraping provides up-to-date shelter availability
What we learned
Multi-Agent Systems Are Powerful But Complex
- Agent coordination requires careful message handling
- State synchronization is critical for distributed systems
- Fallback strategies are essential for production readiness
Document Intelligence Needs Multiple Layers
- Single-tool solutions aren't robust enough
- Combining specialized tools (LlamaParse + Gemini) improves accuracy
- Confidence scoring enables manual review workflows
What's next for CareLink
Mobile App
- Native iOS/Android apps for social workers
- Push notifications for case assignments
- Offline mode for areas with poor connectivity
SMS Integration
- Text updates for patients without smartphones
- Bilingual support (English/Spanish)
- Two-way communication for confirmations
Log in or sign up for Devpost to join the conversation.