Runs detection on a box at the site, sends events rather than video, and produces evidence a dispute can actually be settled with.
It is scoped for factory, warehouse and distribution-hub operators who already own ONVIF or RTSP cameras. We do not replace a camera, and we do not sell you one.
Validation only. Nothing is sold and no performance figure is published until it has been measured on a customer’s own cameras, at that customer’s site. If the measurement does not earn it, the product does not ship. It has no users, no pilot institutions and no customers, so everything below describes what it is specified to do rather than what anyone has yet done with it.
Dock and perimeter losses are discovered weeks later during a count, not at the moment they happen. Finding the relevant footage afterwards takes hours or days. And recording fails silently — a camera stops writing and nobody learns of it until the day someone needs what it should have recorded.
How Nimikh VisionEdge is specified to behave, drawn as a flow. It is a designed workflow, not a recording of anything that has run.
At the site
Cameras the site already owns
ONVIF or RTSP. Nimikh does not replace a camera and does not sell one.
Existing camerasDetection on a box at the site
A detection is filtered by class and validated over time before anyone is told. In shadow-trial mode it runs without raising an alert.
Edge onlyWhat leaves the site
Events, not video
Only events, metadata and short clips ever leave the site. Continuous video does not.
Events onlyEvidence with a chain of custody
Every export carries a hash and a chain-of-custody record. No role holds a permission to delete evidence.
Hash + custodyEach of these is a numbered entry in the requirements document, and every capability further down traces back to one of them. They are described problems, not measured ones, so none of them carries a figure.
Loading-dock and perimeter losses are discovered at stock-take rather than at the moment they happen.
By then the footage window has closed and the loss is unattributable.
Finding the relevant footage across many cameras takes hours or days.
The cost of answering a question exceeds the value of the answer, so the question stops being asked.
Recording fails silently — a disk, a power supply, a camera goes offline.
Nobody notices for weeks, and the failure is discovered on the one day the recording was needed.
There are no figures in this section, and their absence is the point rather than an oversight.
The numbers that circulate about video security — the share of alarm activations that turn out to be false, the reduction a detection system is said to deliver, the number of streams a box will carry — are either measurements of somebody else’s equipment or could not be traced back to a method we were able to check. Printing one here would tell you something accurate about a system that is not this one.
How many streams a box handles depends on resolution, frame rate, how busy the scene is and which accelerator is in it. How many alerts are worth acting on depends on your lighting, your camera angles and what actually moves through the frame at night. Neither is knowable from here.
What replaces the figure is a measurement. Shadow-trial mode is specified for exactly this: detection runs at the site without raising an alert, and what it produces is that site’s own baseline. That number is the only one we will ever quote for your site.
This is a standing commitment rather than a position for the moment. No accuracy, detection-rate, false-alarm or camera-count figure is published for VisionEdge before it has been measured, and that holds after the product ships as much as it holds today.
Written as designed, not as delivered. Nimikh VisionEdge is not built, and each entry below is a specified behaviour a buyer can hold us to once it is — which is the only reason to write it this precisely before it exists.
Detection runs on a box at the site. Only events, metadata and short clips ever leave. Cloud video analytics at USD 20–50 per camera per month is rejected by buyers here, and cloud GPU costs more than the guard it would replace.
A detection is filtered by class and validated over time before anyone is told about it. An operator who is interrupted for nothing stops responding to the alerts that matter.
An export carries a hash and a chain of custody, because the point of the footage is that someone will dispute it.
A camera that has quietly stopped recording raises an alert. This is the failure that makes every other capability worthless.
Almost every number in the specification is a threshold with a consequence attached rather than a figure chosen to impress. These are the ones a buyer would notice, published before the software exists so that they can be held against us rather than discovered afterwards.
A camera or a recording that has quietly stopped is detected and raised. This is the failure that makes every other capability worthless.
A shadow trial in which detection runs and no alert is sent, producing the site’s own baseline before anything is sold against it.
Carries a hash and a chain-of-custody record, because the point of the footage is that somebody will dispute it.
Holds a permission to delete evidence — not an operator, and not the site administrator either.
Design commitments, not results. Not one of these has been measured on a running system, because there is no running system. They are what the software has to do before we would call it finished, and none of them becomes a claim until it has been measured.
Nothing here is an omission, and this section matters at least as much as the one above it. Every entry was decided in the requirements document with a reason attached, and every one of them costs us a feature a competitor will happily claim. A product defined only by what it does is a product whose boundaries you find out about after you have paid.
No independently audited figure exists for this class of system, and a model-quality claim without a measured per-site baseline is a claim about someone else’s cameras. The number comes from your site or it does not get printed.
How many streams a box handles depends on resolution, frame rate, scene complexity and the accelerator. Quoting it in advance is guessing with the customer’s money.
Biometric data is sensitive under the Personal Data Protection Act and export-restricted. It is the highest-liability feature available and there is no verified demand for it at this price point.
We integrate commodity hardware. Capital, supply chain, duties and returns are someone else’s business.
Built in a fixed order, and none of them available today.
The next step for a site is a measurement, not a quotation: detection running in shadow mode on your own cameras, in your own building, producing your own baseline. If that baseline does not earn the product, we will tell you so and stop.