HumanityAI'd — Patrol Sight¶
Download
Your officer looks at a face. They know who it is before the person walks past.¶
Patrol Sight puts face recognition into a security officer's line of sight. The officer wears lightweight AR glasses; what they see streams to a server inside your building, which identifies faces against your own watchlist and returns a colour-coded label into the officer's lens in about 68 milliseconds — hands free, no phone, no radio call, no waiting for the control room.
Why security operations choose it¶
- Recognition where the officer is — not at a gate, not in the control room, not after the encounter.
- Measured, not claimed — ~68 ms glasses-to-label (49–103 ms range), 14–20 recognition frames per second, published stage by stage.
- Your watchlist, your server, your rules — nearly 94,000 people enrolled in the live reference deployment, with a documented path to 13 million. No scraped database, no cloud, no vendor-held face data.
- Sovereign by construction — biometric data never leaves the building, satisfying UAE and Saudi data-residency requirements by architecture rather than by promise.
- A human always decides — every match is presented to an officer to verify. The system informs a decision; it never authorises one. Every search and match is logged and attributable.
Proven, not promised¶
A full production deployment runs today: a 34-service stack on a single GPU, 93,728 enrolled persons, glasses-to-label recognition measured at ~68 ms typical with 0% packet loss, and a watchlist search P95 of 8.5 ms across nearly 2,000 real searches. The same software, deployment tooling and operator console you would receive.
How it works¶
AR glasses stream hardware-encoded H.264 over your dedicated wireless network → your on-premises GPU server detects faces and matches them against your watchlist (ArcFace on TensorRT + GPU FAISS) → a colour-coded category label appears in the officer's lens while the control room sees the same alert live. A two-layer tracker answers 97.6% of frames from cache, which is what makes continuous recognition affordable on one GPU.
What it does not do — stated up front¶
Reliable identification to about 5 metres (6–7 m in good light, 8–10 m in long-range mode) — beyond that, fixed cameras remain the right tool. No automated enforcement action. No emotion or demographic inference. No covert-operation features. We publish the envelope because a security buyer will find it anyway.
Ideal first deployment¶
A high-footfall public realm or transport environment where officers already patrol on foot and a control room already runs facial recognition on fixed cameras. Start with a 6-week, one-site pilot — including an information-security and data-protection gate before any live watchlist is loaded — and carry the watchlist, thresholds and audit history straight into rollout.
Let's scope your pilot. · HumanityAI'd · Artificial Intelligence for Smart Glasses