Inspiration

Millions of older parents live alone while their children work in another city. Many already wear a fitness band that tracks heart rate, oxygen and falls — but nobody watches that data. When something goes wrong, families find out late, or get noisy alerts that cry wolf every time Appa goes for his morning walk.

What it does

CareCircle watches health-band readings and spots possible concerns: a racing heart while resting, low blood oxygen, or a fall. An AI agent on Amazon Bedrock AgentCore then checks the context — the person's usual resting heart rate, their recent activity, the last hour of readings — and decides how serious it is.

Alerts the right person by email (Amazon SES) with one-tap replies: I'm checking, Show trend, All fine now.
Answers questions like "How is Appa now?" using the real readings.
Coordinates the family — when one person says they're checking, the others are told.
Escalates automatically to the next family member if nobody responds.
Ignores false alarms, like a high heart rate during a walk.
Sends a daily summary email written by Amazon Bedrock.

How we built it

Amazon Bedrock AgentCore hosts a Strands agent running Claude Sonnet 4.5, with tools to read vitals, assess incidents, send alerts and schedule escalations.
Amazon SES delivers alerts, follow-ups and daily summaries; AWS End User Messaging (SMS and WhatsApp) is built in as optional channels.
AWS Lambda, API Gateway, DynamoDB, EventBridge, SNS and SSM run the pipeline, all deployed with AWS CDK.
Every agent decision is written to an incident timeline shown on a live dashboard.
A rule-based fallback takes over if the AI is unavailable, so an alert is never lost.

Challenges we ran into

Avoiding alert fatigue (exercise vs. a real problem), making replies work without logins (signed one-tap links), and keeping AI decisions auditable and safe for a health use case.

Accomplishments that we're proud of

A full, working loop — detection, AI assessment, alert, family reply, coordination and escalation — running end to end on AWS, with every step visible.

What we learned

Agents are most useful when they make judgement calls (how serious? who to tell?) while safety-critical actions stay deterministic.

What's next

Real wearable integrations (Health Connect, Fitbit), WhatsApp and RCS as primary channels, long-term per-person baselines with AgentCore Memory, and voice-call escalation.