Incident.response

Inspiration

You love Uncle Rodger, but you can't always be there for him. If he falls when he's home alone, he could lie there for hours before anyone knows. Falls are one of the biggest risks for older people living alone, and the usual fixes don't work well. People forget to wear pendants, can't reach panic buttons after a fall, and many say "I'm fine" when they aren't. We wanted something that needs nothing from him, treats him with respect, and still gets family there fast when it matters.

What it does

Detects a fall with an ordinary camera and saves a short clip.
Asks Uncle Rodger out loud: "Are you okay?" If he's fine, it closes quietly.
Escalates if he asks for help, is unclear, or stays silent. Family gets a WhatsApp alert with the clip and an AI phone call, one contact after another until someone responds.
Keeps him informed by telling him out loud when help is coming.

How we built it

Vision: OpenCV + MediaPipe pose tracking with a fall state machine
Backend: FastAPI + SQLite incident engine with timed escalation
Voice: ElevenLabs agents for the resident check-in and caregiver calls (via Twilio)
Messaging: WhatsApp alerts, clips and replies via neonize
Every service has a mock version, so the whole system is tested offline (742 tests)

Challenges we ran into

Getting the video uploaded before the WhatsApp alert goes out
Making sure silence or a mumbled answer never counts as "I'm fine"
Hardware surprises: the wrong default camera, and Bluetooth audio crashes

Accomplishments that we're proud of

A full loop: camera → fall → voice check-in → resolved quietly or family alerted with video
False alarms stay private, without panicking the whole family

What we learned

How much getting quick, accurate alerts about dangerous situations matters, because the sooner the right people know, the sooner someone can step in and help.