Data Hatch
Face recognition and anti-spoof liveness running fully on-device across ESA Automation’s industrial fleet: the perception layer for agents that gate machine controls, triage failed checks, and monitor model drift across sites.
Operators authenticated into ESA’s HMIs and industrial PCs with typed usernames and passwords, with gloves on, in dust and vibration, across multiple machines per shift. In practice, credentials got shared, sessions stayed open, and there was no reliable record of who was actually at the machine.
Cloud face recognition was a non-starter: factory networks are segregated, latency matters at the machine, and biometric checks had to stay on site. Everything had to run on the device itself, across the two hardware architectures already in the field.
A detection model finds and locks the operator’s face in the device camera stream, in factory light conditions.
An embedding model matches the face against the enrolled operator registry, entirely on the device.
A liveness layer rejects photos, screens, and replays before any match is trusted.
Grant or deny in milliseconds. Machine controls unlock only for a live, recognised operator.
Agents triage denied checks: unknown faces, spoof attempts, and repeated failures are escalated with context.
Model performance is monitored across the fleet, so lighting changes, new PPE, or camera wear surface before they become failures.
Nothing leaves the device: no images, no calls.
Every inference runs on the device. Nothing leaves the site.
One stack across ARM and Intel IoT hardware.
Detection, recognition, and liveness anti-spoof.
Operators authenticate hands-free, gloves on.
Millisecond decisions, no network dependency.
The foundation for agents that gate controls and triage failures.