Presence

Attendance systems often stop at forms or QR codes, which can be shared around. I wanted to test a more accountable flow where the check-in is tied to a real face, a live person, and the employee's assigned shift.

I split the app into a public kiosk and an admin workspace. Face descriptors are extracted in the browser, matched server-side through Supabase pgvector RPCs, then written through protected server routes after liveness and shift checks pass.

Next.js keeps kiosk, admin, and server routes in one codebase. Supabase gives Auth, Postgres, Realtime, and pgvector; TensorFlow plus face-api.js handles recognition without sending raw images to a third-party model.

A public MVP demonstrating browser face embeddings, server-side pgvector matching, shift checks, realtime monitoring, and PDF reports. Head-movement liveness is demonstration-grade and has not been validated as production anti-spoofing.

I learned that face recognition is mostly an operations problem: calibration, duplicate checks, security boundaries, and clear feedback matter as much as the model itself.