Lumi AI

Inspiration

My grandmother's friend has dementia and lives in an understaffed senior home—a reality shared by millions. U.S. nursing homes often average a 1:10 caregiver-to-resident ratio, sometimes reaching 1:15. When we saw the community-centered track, the problem felt deeply personal. People who need urgent help deserve faster, smarter, and more dependable support.

What it does

Lumi is an always-on AI care companion for seniors with dementia. It uses facial recognition to identify visitors, describes incoming people, flags unknown faces and asks for their names, detects falls in real time, listens for distressed calls, gives medication reminders, and sends automatic SMS alerts to caregivers and emergency contacts.

How we built it

We built Lumi with a Raspberry Pi, webcam, and an iPhone interface. The camera captures live video, matches faces against a Supabase database, and flags unknown visitors. The mobile app was built in Swift, Firebase handles authentication, Gemini adds image understanding, and a validation layer checks AI responses before storage.

Challenges we ran into

The hardest early challenge was finding an idea that actually mattered, so we turned to our own lives. Technically, Base64 image data would not cleanly convert into the JPEG format we needed. We also ran into Raspberry Pi CPU issues, which forced us to shift part of the system into a locally hosted setup.

Accomplishments that we're proud of

We are proud of building a fall detector that is genuinely useful. We were also able to get our front-end pivoted from our current idea with a time crunch.Both were difficult to implement, but they made the product feel real and impactful.

What we learned

We learned that hardware adds an entirely different layer of complexity compared with software alone. We also learned that the strongest ideas usually come from personal experience, not abstract brainstorming. Most importantly, we learned that validating AI outputs early saves huge amounts of time and prevents unreliable behavior later.

Built With

  • chatgpt
  • claude
  • codex
  • cursor
  • gcp
  • gemini
  • limelight
  • raspberrypi
  • windsurf
Share this project:

Updates