ContinuumCare

Inspiration

Healthcare is personal for us. We’ve seen families struggle with unreadable discharge papers and caregivers burn out while monitoring loved ones with dementia. With 1 in 5 patients readmitted within 30 days, we built ContinuumCare to be the safety net that hospitals currently lack. We created this for every family that has felt abandoned the moment the hospital doors closed.

What it does

ContinuumCare is a multi-agent AI platform that ensures patients are never alone after discharge:

AI Pipeline: Five specialized agents (Orchestrator, Groq-powered Clinical Reasoning, Content, and Safety) process cases in parallel.

Human-in-the-Loop: Before any recommendation reaches a patient, the system pauses for a mandatory clinician approval gate.

Accessible Output: Generates clear care plans read aloud via ElevenLabs for patients with low literacy or vision impairment.

Active Monitoring: Passive sensors detect falls (accelerometer), distress (Web Speech API), and facial pain (TensorFlow.js), while a Galaxy Watch 7 streams real-time vitals to a clinician dashboard and a caregiver PWA.

How we built it

Backend: FastAPI on Railway using Anthropic (Claude) for orchestration and Groq (Llama 3.3) for reasoning. We used asyncio.Event to halt the pipeline in memory for human approval.

Frontend: Next.js 14 on Vercel. Features a real-time communication log and a mobile PWA for caregivers.

Sensors: No-install Android web apps utilizing DeviceMotion, Web Speech, and Samsung Health SDK.

Integration: A strict, pre-defined API contract allowed three developers to work in parallel and integrate seamlessly.

Challenges we ran into

CORS & Deployment: Faced preflight blocks between Vercel and Railway. We bypassed this using a Next.js rewrite proxy to route all API calls through a single origin.

SSE Streaming: Maintaining persistent Server-Sent Events through reverse proxies required specific header configurations to prevent premature termination.

Hardware: The Galaxy Watch 7 SDK was more complex than anticipated, requiring developer mode and companion app pairing under intense time pressure.

Accomplishments that we're proud of

The Approval Gate: Our system is the only one that prioritizes safety by refusing to act autonomously in a medical context.

Real Data: We integrated a real heartbeat from a Galaxy Watch 7, not a simulation.

Impact: Solving accessibility for elderly patients by replacing jargon-heavy paper with clear, spoken instructions.

What we learned

Human Coordination: Agreeing on a "sacred" API contract before coding is the best way to manage multi-person AI development.

Emotional Resonance: The most effective demos aren't just technically complex—they are the ones that solve a visible human problem.

What's next for ContinuumCare

EHR Integration: Connecting directly to Epic/Cerner workflows.

DementiaOS: Longitudinal pattern detection to identify cognitive decline over time.

Audit Trails: Clinician credentialing for the approval gate.

Multilingual Support: Serving underserved communities in their native languages.

Built With

• anthropic-claude
• content-agent
• devicemotion-api
• elevenlabs
• fastapi
• groq
• llama-3.3
• mongodb-atlas
• motor
• next.js
• ngrok
• python
• python-dotenv
• railway
• react
• samsung-health-sdk
• sse-starlette
• tailwind-css
• tensorflow.js
• typescript
• vercel
• web-speech-api