MemoLens

Inspiration We built MemoLens to reduce everyday confusion for dementia patients and reduce anxiety for caregivers. The goal was simple: make memory support work in real time, not after the fact.

What it does MemoLens recognizes familiar people, tracks important items (like keys/medication), triggers contextual reminders, and answers memory questions such as “Where are my keys?” Caregivers see live updates in a dashboard/timeline without manual refresh.

How we built it We built a full pipeline: Pi/webcam stream -> FastAPI backend -> CV + reminder logic -> Postgres event/state store -> live WebSocket feed -> Next.js caregiver/patient UI. The backend handles auth, patient data, reminders, notes, events, and query routing. The frontend provides caregiver management pages and simplified patient views.

Challenges we ran into Setting up reliable local infra (Postgres/Redis/env vars) across machines Handling native CV dependency issues (dlib/face_recognition) in dev environments Ensuring real-time updates were stable and not dependent on manual refresh Making demo flows reproducible end-to-end

What we’re proud of Real-time caregiver feed and timeline working end-to-end Context-aware reminder triggers (time/person/location/object + daily notes) End-to-end smoke test and deployment check scripts Demo rehearsal tooling that validates the exact presentation flow

What we learned We learned that reliability and integration matter as much as model accuracy. Getting the full system to work consistently (stream, inference, DB writes, and live UI updates) was the biggest win.

User attachment (No content) Python, TypeScript, FastAPI, Next.js, React, SQLAlchemy, Alembic, PostgreSQL, Redis, Celery, WebSockets, OpenCV, Ultralytics YOLOv8, PyTorch/TorchVision, NumPy, Pillow, face_recognition (optional), Raspberry Pi + USB webcam, Docker, Docker Compose, Render, Vercel, Supabase Storage.