Camera intelligence · events, not footage · on-prem

Rules you write in words.

Point it at the cameras you already have and write the rule in a sentence — “tell me when a pallet blocks the fire exit.” When it happens you get an event: the time, a line explaining why, and a still with faces and plates blurred. Not a recording. It runs on a box on your site, and nothing leaves the building.

Blocked exitSpillNo helmetAfter-hoursWrong bayLeft running
Book a scoping call See the four-week proof
Why it is different

Three things the usual video analytics does not do

Anyone can bolt a model onto a camera. The difference is in the three things below — and the third is the one a serious buyer asks about first.

A rule is a sentence

A new rule does not need labelled footage and a trained detector. It needs a sentence. You add or change a rule the same afternoon your process changes — not in the next project.

We keep the event, not the footage

The box stores the moment, the reason and a still with faces and plates blurred. Not a recording of your site. That is a selling point and a legal stance at once: you hold evidence of an event, not an archive of people.

We tell you how often it is wrong

The false-alarm rate per rule, measured on your site before you roll out. The same evaluation discipline we apply to text models, turned on pictures. Nobody else hands you that number.

How it works

A cheap gate first, the model only on what matters

Running a vision-language model on every frame is not affordable, so it never sees every frame. A cheap gate watches for motion and scene change; only the candidates reach the model, which judges them against your rules; a reasoning step suppresses repeats and decides what to escalate.

CameraEvery frame, all day.
GateMotion, scene change. Most stop here.
ModelOnly the candidates, judged against your rules.
EventWhat actually matters.

Cameras

Your existing camerasRTSP · no new wiring
Cheap gatemotion · scene change

The box, on your site

Vision-language modelreads the candidate against your rules
is the rule met?whyconfidencededupeescalate
Nothing leaves the building

Event

Time and the rule that fired
A line of reasoning
A still with faces and plates blurred
Alert: app, e-mail, Telegram
The gate is the economicsThe model runs on a handful of candidates per camera, not on a video stream, so the cost per camera stays sane.
The reasoning step is the hard partA raw “is this notable?” fires too often. Suppressing repeats and escalating only what matters is what the competition does not have.
Search by exampleUpload one picture — a vehicle, a piece of equipment — and get every past event with it, without training a detector.
On your termsThe box runs on your site or on hardware you rent in the EU. The events are yours.
What you can ask

Rules written the way you would say them

Each of these is one sentence to the system. You keep the ones that matter to you and change them whenever the site does.

safetyTell me when a pallet or anything else blocks a fire exit.
PPETell me when someone in the yellow zone has no helmet or vest.
spillTell me when there is liquid on the floor in the aisle.
accessTell me when a person is in the plant room after working hours.
logisticsTell me when a truck is in the wrong loading bay.
equipmentTell me when a machine is left running with nobody present.
Event, not footage

We hold evidence of an event, not an archive of people

Monitoring a workplace is specifically regulated in Slovenia (ZVOP-2) and under the GDPR, and an archive of recordings is the heavy part. Perception Watch sidesteps it: it keeps the event and a still with faces and licence plates blurred, and the original frame can be restored only with a key held by an authorised person. That is the same reversible pseudonymisation we built for documents, turned on pictures.

09:42:17face ×2 · blurred
09:42:17plate ×1 · blurred
09:42:17recording · not kept
09:42:18event + still · kept
10:03:40face restore ×1 · authorised key · logged
No recording keptThe stream is watched, not stored. What persists is the event and a redacted still.
Faces and plates blurredPersonal data is masked on the way in; restoration needs an authorised key, and every restore is logged.
Reversible, not destroyedWhen an incident genuinely needs the original, an authorised person can recover it — the default is masked, the exception is logged.
A stance you can defendTo an inspector or a works council, “we keep the event, not the footage” is the difference between “interesting” and “allowed”.
Where it fits

Many cameras, a few rules that keep changing

The easiest first customer is not a factory with a built-in VMS and a two-year procurement cycle. It is somewhere with a lot of cameras and a handful of rules that change often.

A good fit

  • Utilities and municipal yards
  • Construction sites and depots
  • Logistics terminals and warehouses
  • Building and facility management
  • Agriculture and outdoor sites

Not what this is for

  • Exact counting of items through a point
  • Precise localisation or measurement
  • Reading a licence plate character by character
  • Anything where a classic detector, trained once, is simply the right tool — we will say so
The proof · early access

Four weeks, one site, three rules, a number

Perception Watch is in early access and we are looking for the first sites. We do not ask you to believe a brochure — we measure on your cameras.

1 · Scoping callYou describe the site and the three rules that would matter most. We say honestly whether they suit this tool or a classic detector.
2 · The box goes inWe connect to the cameras you already have. No new wiring, no cloud account, nothing leaves the building.
3 · We measureFour weeks later you have the false-alarm rate per rule on your own footage — the number that decides whether it is worth rolling out.
4 · You decideIf the number works, we roll out more cameras and more rules. If it does not, you learned that cheaply and early.
Book a scoping call ↗
FAQ

Questions buyers ask

How is this different from the video analytics my camera vendor sells?

Two ways. A new rule is a sentence, not a trained detector with labelled data, so you can add 'tell me when a pallet blocks the fire exit' the same afternoon. And we keep the event, not the footage: the box stores the moment, a line of reasoning and a still with faces and plates blurred, not a recording of your site.

Does the video leave the building?

No. The box runs on your site. Nothing is sent to a cloud. What it stores is the event and a redacted still; the original frame can be restored only with a key held by an authorised person.

Can it count things or read a licence plate exactly?

For exact counting and precise localisation a classic detector is better, and we will tell you so. Perception Watch is for 'tell me when something happens that I did not list in advance' — a blocked exit, a spill, a person in a zone after hours.

How do I know it is not crying wolf?

That is the part we measure. A four-week paid proof on one of your sites with three of your rules gives you the false-alarm rate per rule before you roll anything out. If it is above a workable threshold, you find out early and cheaply.

What do I need on site?

The cameras you already have and one box we supply or you rent. No new wiring, no cloud account. Rules are written in a sentence and edited whenever your process changes.

Talk to us ↗