Permission
The browser grants camera access only after a direct user gesture over HTTPS.
Case 01 / Edge AI
A front door
with boundaries.
The proposition
What if intelligent security could recognise a resident without turning the home into a surveillance platform?
The prototype uses explicit enrollment and one-to-one verification. It answers one narrow question—“does this live image match the person who enrolled in this session?”—instead of searching a database for identity.
Interactive demonstration
Verify yourself.
Nothing more.
This demo activates only after you choose to continue. Frames are sent to the private Raspberry Pi service for processing, discarded immediately, and never added to an identity database. Your temporary face embedding expires after ten minutes or when you press Delete.
Camera ready
Enroll this session
Look straight at the camera in even light. Exactly one face must be visible.
Run live verification
After enrollment, start comparison. One frame is checked at a controlled interval.
System design
Realtime without the cloud.
Boundary
PHP validates the session, action, image type and payload size before proxying internally.
Inference
Python uses YuNet to detect one face and SFace to create and compare an embedding.
Decision
The browser receives a match score—not a face image or a reusable identity record.
What makes it responsible
Consent before capture. No camera request occurs on page load.
Verification, not identification. The system compares only against the current visitor’s temporary enrollment.
Data minimisation. Frames are decoded in memory and discarded; embeddings expire automatically.
Defense in depth. Facial verification should complement a physical key, PIN or trusted-device factor—not replace them.
Prototype limitation. This demonstration does not implement presentation-attack or liveness detection and must not directly unlock a production door.