We started by putting cameras up. We carry on making them understand.

In 2006 the job was putting cameras up and making them work. It was done well and it still is: banks, factories, transport, events. None of that has stopped being needed, and anyone who says artificial intelligence replaces it has never climbed a column.

But over the years an uncomfortable thing became obvious. Every new installation added screens to a room where the number of people watching did not grow. More and more was recorded and less and less was seen, and video became the thing you consult after something has happened instead of the thing that warns you while it is happening.

Infinity Neural was set up in 2011 to solve that, and not to sell something else: the same camera that was already there starts to understand what is in front of it. It is the same idea as always, finished.

The usual alternative is «industry leaders», «excellence», «comprehensive solutions». Phrases that work for anybody and that therefore say nothing about anybody. We would rather say what we are good at and what we are not, which is a conversation you can check.

We are good at looking at an installation and knowing what can be got out of it. At putting a camera where it has to be and not where it is convenient. At tuning a detection for a particular scene until it stops failing. And at answering for both things at once, which is the rare part: hardly anyone has the person who decides the height of the column and the person who decides what gets understood from that image under the same roof.

And there is one thing measured by somebody from outside, which is the only serious way of saying that something works: people counting is approved by the Centro Español de Metrología. It is not us saying it, it is a body that does not sell the product.

We do not promise a hundred per cent accuracy. A system that looks at images gets things wrong, and anyone who says otherwise has never put one into production. What you can do is measure how often it gets things wrong and under what conditions, and that we do show you.

We do not promise zero false alarms. We promise to work at keeping them few, because a system that alerts five times a night because of a branch stops being a security system by the third week: whoever is on duty learns to ignore it.

We do not promise that the cameras you have will do. Sometimes they will. Sometimes the answer is that the camera is not good enough for what you want and it has to be moved, and we say that at the site visit and not on the day of the incident.

And we do not promise generic timescales or fixed catalogue prices. An installation of twelve cameras and one of six hundred are nothing alike, and giving a number before looking is a polite way of lying.

A security system is not bought: it is maintained. On installation day everything works. What decides whether it is still useful three years later is whether somebody inspects it, whether the alarms are tested, whether the cameras that have stopped seeing properly get cleaned and repositioned, and whether the rules are updated when the operation changes. A supplier cannot hold that up on their own.

That is why the projects that turn out well have a person with a name inside them: the head of security, of operations, of the plant. Somebody who decides what is watched, who gets alerted and what changes when the work changes. The ones that fade out are the ones without that person, and they always fade out the same way: nobody looks at the alerts, somebody silences them, and the following year the system is an expense nobody defends.

And there is one conversation we know how it ends: the one that starts with a list of cameras and the question of what each one comes to. There the honest thing is to say that we are not the cheapest and we are not going to be, because half of what we charge is the project done beforehand and the maintenance done afterwards, and that does not show up in the price per camera.

It is not arrogance. It is that a system installed for somebody who does not want it switches itself off, and then both of us have lost.

Almost everybody buys those two things separately, and that is where the problem starts. One company puts the camera in, another puts the software in, and the day something does not see what it was supposed to see there are two suppliers in the same meeting explaining that the fault is the other one's. It is not bad faith: it is that neither of them can fix the other's half.

Here the person who decides the height of the column and the person who decides what gets understood from that image work in the same house and talk to each other before the wall is drilled. That changes the order of the questions: whether that camera is going to have to read a number plate one day is known at the site visit, not two years later with the work already done.

WestPoint does electronic security end to end: systems engineering, camera systems, access control, intruder detection and fire detection. It designs, installs and maintains, with 24 × 7 support, in companies, banking, transport and factories. And it puts up and takes down the temporary security for large events, which is another trade with different deadlines.

Infinity Neural makes those cameras understand what they see, and it does it in two ways. Made to measure, studying your scene and tuning the detection for your particular case, which is the greatest accuracy possible in THAT case. And with IRIS Neural, a system that covers the whole installation and works with the internet cable unplugged.

Hardly any real installation has only one of the two. The normal thing is IRIS covering all the cameras and two or three made-to-measure detections where it really matters.

It starts by looking. One of your cameras, whichever you like: where it is, what angle it has, what can be seen from there at six in the evening and at six in the morning. That is where the first thing we say comes from, and sometimes it is that it cannot be done, and that is an answer too.

If it can be done, it is done small before it is done big: a few cameras and one or two rules, checking that it detects what it is meant to. Nobody has to believe anything. Then it grows from there.

And somebody stays. Maintenance is not a line in the quotation: it is what makes the system still useful in three years for what it was bought for.

We do not promise a hundred per cent accuracy, or zero false alarms, or compatibility with everything. A system that looks at images gets things wrong, and anyone who says otherwise has never put one into production.

What we do is say what can be done today, what cannot, and what would have to be looked at before answering. When something is measured by a third party —like people counting with the Centro Español de Metrología— we show it. When it is not, we say that too.