Vertical
Video surveillance
Systems that have to produce usable evidence, at night, at a remote site and under audit. The camera is the easy part.
Video surveillance market, scope South America, per Mordor Intelligence. Public figure compiled in 2026. This is market context, not this company's results.
Why it is bought differently
The buyer is not comparing cameras
They are buying the ability to reconstruct what happened, to prove it afterwards, and to still have a working system next year.
Most tenders in the region are written in camera counts and sensor resolution. Both are easy to compare and neither predicts whether the system will do its job. A system is lost on effective coverage, on bandwidth, on retention and on continuity of operation, and none of those four appear on a product sheet.
The result is familiar: systems that meet the tender to the letter and identify nobody, archives overwritten before anyone asks for the footage, and links that cannot carry the stream the design assumed.
So the work starts by writing down what has to be seen, at what distance, and for how long it has to remain recoverable. Everything else follows from that.
Coverage
Detect, observe, recognise and identify are four different things
The international standard for video surveillance systems defines those four levels in pixels per metre on the target, not in sensor megapixels. It is the difference between knowing someone is there, being able to say what they are doing, being able to say it is the same person seen on another camera, and being able to say who they are. A tender that asks for identification and specifies detection optics is not met by adding resolution.
Focal length, camera count and position all follow from that level. A panoramic camera covering an entire plaza meets detection and meets identification nowhere in that plaza. It usually takes a second fixed camera aimed at the choke point, which is where identification is possible at all.
Lighting conditions decide the rest. Backlight at an entrance, rain, fog and moving vegetation degrade the image exactly when it is needed. That is where dynamic range, real sensor sensitivity and the choice between infrared and white supplementary lighting carry weight. That last choice is not technically neutral: white light allows colour and changes how the installation is perceived in an inhabited space.
And one condition is routinely overlooked and voids the evidence: time. Without common time synchronisation, footage from two cameras cannot be correlated and an access event does not line up with its image.
Bandwidth and retention
Storage is sized on the worst case, not the average
Required storage is the bit rate per camera, multiplied by the number of cameras and by the seconds of retention required. The arithmetic is trivial. The error is always in the first factor.
The bit rate a manufacturer publishes is measured on a static scene. In service, compression works on the difference between frames, so everything that moves pushes it upward: wind in vegetation, rain, traffic, a flag. An outdoor scene at night in rain can multiply the consumption of the same camera in the same configuration. If the calculation used the catalogue number, the archive shortens by itself and nobody notices until it is needed.
Codecs with regions of interest reduce that consumption by lowering quality outside the defined area. They work, and they degrade what the analytics can see in that zone. They are used deliberately, not by default.
Retention is an obligation, not a preference: it is set by the customer's policy or by the rule that applies to them, and it is written into the tender before the storage is chosen. On top of that come the fault tolerance of the disk array, the rebuild time after a replacement, and whether there is edge recording so a network outage does not leave a hole in the archive.
The link is the other half. A remote site on a shared radio link will not carry the continuous stream of all its cameras: there, recording happens at the edge and transmission is on demand or on event. That is an architecture decision, and it is taken before any camera is bought.
AI analytics
Classification is what turns an alarm into a notice
Motion detection has been available for decades and produces alarms from rain, insects, shadows and branches. What changed is classification: a vision model distinguishes person, vehicle, vehicle type and animal, and lets the rule be written in terms of the event that matters rather than the pixel that changed. A line-crossing rule applied only to people removes most of the night-time noise on a perimeter at once.
On top of that sit attribute search, which is what makes hours of archive reviewable in minutes, counting and occupancy, and detection of specific behaviours such as loitering, abandoned objects or wrong-way movement. Each has its own error rate and is tuned on site, against the real scene.
Where the model runs is an architecture decision, not a brand one. On the camera it cuts traffic and does not depend on the link, but it is tied to the model the manufacturer shipped. On a server it allows heavier models, permits reprocessing the archive when a new requirement appears, and concentrates cost in accelerators that have to be sized and cooled. Most projects end in a combination, and it is worth deciding it in writing.
The metric we ask for is not a laboratory accuracy percentage but false positives per camera per night, measured on site, in the worst weather. A percentage with no defined scene cannot be verified; a count of false alarms can, and it is what decides whether the operator is still attending the system three months in.
Two categories are specified separately. Licence plate reading needs dedicated optics, shutter and lighting, and is not solved by enabling a feature on a general-purpose camera. Facial recognition, beyond the technical side, is a category with personal data protection implications: it is specified with the customer's legal function, with a stated purpose, or it is not specified at all.
Portfolio
What we specify and integrate
Camera, optics and lighting
Fixed, dome, multisensor and thermal cameras, with optics chosen by coverage level rather than by catalogue. It includes supplementary lighting and power over the network, which governs the switch power budget and outdoor behaviour.
Management platform and recording
Video management, recording, redundancy and export with chain of custody. A multi-brand deployment rests on interoperability profiles; it is worth knowing in advance which functions are lost outside each manufacturer's native driver.
AI video analytics
Object detection and classification, rules by zone and by behaviour, attribute search and counting. At the edge, on a server or combined, decided by available bandwidth and by the need to reprocess the archive.
Storage and video network
Array and segmented network sizing for video traffic, with defined fault tolerance and rebuild time. It is the part of the system bought once and paid for across the whole retention cycle.
Access control and credentialing
Credentials, turnstiles and barriers, visitor management, and integration with video so every access event has an image attached and a synchronised timestamp. Without that integration, the later audit is done by hand, which means it is not done.
Control room
Video wall, event management, response procedure and communication with field staff. It is what turns a detection into a response, and it is usually the first thing cut from a budget.
How we work
How we enter a video project
- Site survey Light and weather conditions, distances, power available at a pole or remote point, and the real capacity of the link. On perimeter work this decides the technology before any brand preference.
- Coverage design by level What has to be seen and at what distance, written by level: detection, observation, recognition or identification. Writing it that way heads off the argument at handover.
- Network and storage sizing Bit rate per camera on the real scene, required retention, fault tolerance and backup policy. This is where much of the work specified purely by camera count falls over.
- Analytics tuning on site Rules configured against the real scene and false positives measured per camera per night over an agreed period. Analytics are not delivered on installation day.
- Support and second line Local support with escalation to the manufacturer. At a remote site, the gap between a spare in 48 hours and one in six weeks is the gap between a system running and a system off.
National and regional government · Mining · Energy · Ports and critical infrastructure · Corporate · Retail