The analytics run on the camera. The footage never leaves it.
An on-device computer-vision SDK for AI camera makers and the sites that deploy them — perimeter and border security, checkpoints, critical infrastructure and factory floors. It works with no internet at all, and per-unit licensing and metered billing are built into the runtime.
No sign-up, no upload. The demo below is the SDK's own detection code running on your machine.
Not a video. The SDK's own detection code, in your browser.
A direct port of the reference backend — same algorithms, same constants, same tracker and zone logic. Point it at a synthetic scene or your own camera; either way the pixels never leave your machine.
Modules
A module the selected SKU does not entitle is struck through and cannot be
switched on — the runtime refuses it, exactly as
xcaicx_engine_create does.
Controls
Event stream
This is the reference backend, and it is not production accuracy. Contrast, edge and colour heuristics with no learned model behind them — built so an integrator can validate frame plumbing, licensing and events before a model reaches them. On a real site it will miss things and invent things. Production deployments ship real models validated on your own footage.
Four primitives, and most of physical security falls out of them
A tripwire is a virtual fence. A polygon is a restricted area. A dwell timer is loitering. A tracked id is one incident instead of forty.
Perimeter intrusion
A directed tripwire fires LINE_CROSS the moment a track crosses
it — and tells you which way, so a patrol leaving does not raise what a breach
raises.
Restricted areas & loitering
A zone with a dwell timer separates walking through from stopping and staying. Dwell fires once per visit, so the operator gets one alert rather than a stream of them.
Border segment monitoring
Crossings reported with direction, so ingress and egress separate at the camera. Tracks carry ids, so a group crossing together is a countable set rather than one blob.
Vehicle checkpoints
Plate reading at the lane, on the camera, at the barrier — no round trip to a datacentre in the decision path. Add a zone per lane for occupancy and throughput from the same frames.
PPE compliance
Helmet and hi-vis checks on each detected person. It reports the absence, because the actionable event on a floor is the violation.
Surface & defect inspection
Tile statistics against the frame's own robust median, so a defect cannot inflate the statistics used to find it. Deliberately conservative — a false reject stops production.
Dock doors & bays
Directed crossings give arrivals and departures separately; a zone on the bay floor gives how long it stayed occupied. Two calls, and the yard stops being the part nobody has numbers for.
People flow, without identifying anyone
Counts, directions and dwell from anonymous tracks that exist only while the person is in frame. No face recognition and no re-identification — usually the fastest route through a privacy review.
Built for the site that will never give you a network
Most vision vendors treat the air gap as an exception to work around. Here it is a supported path — because the customers who need it are defence, border, utilities and anything with an OT boundary.
Licensed without calling home
Tokens are minted offline and bound to each unit's silicon. The camera verifies against a key compiled into the firmware. No activation server in the loop, and no phone-home to fail.
Outages delay billing, not operations
Usage spools to disk and ships when the link returns. If a licence lapses mid-shift, analytics stop and the video path keeps recording — a camera that stops recording is a warranty claim.
Verifiable, not just promised
Build with TLS off and the HTTP client is not linked into the binary at all. Your
customer's security team can confirm that with ldd rather than
trusting a datasheet.
What the camera sends when it is online, in full: a one-time activation carrying its hardware fingerprint, then a heartbeat carrying counts — inferences, frames, stream-seconds, events. No images, no crops, no identities. There is also no facial recognition, no biometrics and no re-identification in the SDK — a product boundary, not a roadmap gap. It is not a safety interlock and not a targeting input, and we will say so in a bid document as readily as here. Details in the security documentation.
Every camera books itself. Try to break it.
An OEM buys a batch; each camera activates once and consumes one seat, with reflashes free. Licences are Ed25519-signed and bound to the unit's silicon — the lab below mints a real token in your browser and verifies it with the same rules the camera applies.
Token
Verifier verdict
Key is generated in your browser and discarded
when you close the tab. The production root key never leaves the licensing
server; its public half is compiled into the SDK.
5 units remaining on this batch.
Server responses
Offline licensing is deterrence, not DRM. An attacker with root and a disassembler can patch out any check that runs on hardware they control. What this buys: casual copying is dead, entitlements cannot be widened without forging Ed25519, and seat counts are server-side where a device cannot inflate them.
Tell us the board and the site
Evaluation licences cover the full module set. Compose the brief here and skip two rounds of email.
What buyers ask first
Does any video or imagery leave the camera?
No. Inference runs on the unit, and a camera configured offline never opens a socket. What optionally leaves is an activation on first boot and a heartbeat carrying counts — no frames, no crops, no identities.
Can it run completely air-gapped?
Yes, as a documented path rather than a workaround. Licences are minted offline and bound to each unit's fingerprint; the verification key is compiled into the firmware. Build with TLS disabled and the HTTP client is not in the binary at all.
Does it do facial recognition?
No, and it will not. It detects people, vehicles and objects, tracks them while in frame, and reads number plates. No biometric matching, no cross-camera re-identification. For most buyers that materially shortens the privacy review.
What accuracy should we expect?
That depends on your camera, lens, mounting height, lighting and test set, which is why we will not quote a number before seeing your footage. What ships in the box is a deterministic heuristic for bring-up; production accuracy comes from real models validated on your own data.
What does it cost?
Per programme. There are two revenue lines — a per-unit royalty or per-camera subscription, and metered inference over an included allowance. Tell us volume, modules and connectivity and you will get real numbers rather than a list price to negotiate down from.