Familiar AI

Inspiration

Over 55 million people worldwide live with dementia. One of the most painful daily challenges they face isn't dramatic, it's quiet. It's the moment a familiar face walks through the door and they can't remember who it is. That confusion creates stress, embarrassment, and erodes independence. We built Familiar AI because we believed that technology could give that moment of recognition back calmly, privately, and without friction.

What it does

Familiar AI is a camera-based assistive memory system for people with dementia. A Raspberry Pi camera detects faces in real time and compares them against stored identities. When a match is found, a Memory Card instantly appears showing the person's name, relationship, and when they were last seen with an optional spoken reminder via text-to-speech. When an unknown face appears, the system calmly asks if the user wants to save who this person is. The user can speak naturally, "This is John, my neighbor, he visits on Tuesdays" and the system builds a Memory Card automatically. No forms, no menus, no complexity.

How we built it

Familiar AI is built across four layers. A Raspberry Pi with a camera handles physical sensing in the environment. An on-device perception layer running Python and OpenCV detects and isolates faces, generates real face embeddings, and compares them against stored identities using a tuned similarity threshold. Supabase serves as our backend — storing Memory Cards, recognition events, and user data with Row Level Security policies. A React and Tailwind frontend gives the user a clean, accessible dashboard to view and manage their memory cards. For the AI pipelines, we integrated ElevenLabs for both speech transcription and text-to-speech, making the experience fully accessible without typing. Featherless.ai handles serverless inference — taking raw spoken transcripts and automatically extracting structured fields like name, relationship, and visit pattern to build Memory Cards automatically. We used Opennote as our collaborative planning journal to document schema changes and architecture decisions throughout the build. Our biggest engineering challenge was tuning the face similarity threshold — balancing it so the same person is always recognized as the same person across different lighting conditions and angles, without merging two different people into one identity.

Challenges we ran into

Setting up the Raspberry Pi headlessly and debugging the camera field of view was harder than expected. We also spent significant time troubleshooting Supabase Row Level Security policies to get the Pi, backend, and frontend all communicating correctly. Mid-build, we migrated our entire data schema from separate 'people' and 'events' tables to a unified 'cards' structure to improve retrieval speed. Getting the face embedding threshold right took multiple iterations across different lighting and angles.

Accomplishments that we're proud of

We built a fully working end-to-end hardware and software system in a single hackathon. The Pi camera detects a face, the system recognizes it, and a Memory Card appears on screen with a spoken reminder — all in under a second. We're especially proud that the primary user is the patient themselves, not a caregiver. It's calm, dignified, and actually useful in daily life

What we learned

We learned how to integrate physical hardware with a cloud backend under time pressure. We deepened our understanding of face embeddings, similarity thresholds, and RLS policies in Supabase. We also learned that the best assistive technology is the kind that gets out of the way — simple, fast, and built for the person who needs it most.

What's next for Familiar AI

We want to add caregiver and family alerts so loved ones are notified of recognition events. We're also exploring a photo memory album feature, doctor portal access, and expanding to mobile so the system isn't limited to a fixed Pi camera. Long term, we believe Familiar AI could be deployed in care homes as a standard assistive device.

Built With

• bash
• javascript
• opencv
• python
• supabase