From cameras nobody watches to an alert in seconds.

1. Weapon detection

   The AI didn't know a real room.

   The problem

   Our starting weapon model had learned from public photos of guns and knives in good light. On our office webcam it found a kitchen knife in only 3 to 6 frames in a hundred, and a cleaver in none. On a dual-lens camera that stacks its two views in one picture, it found a rifle in 0 of 84 test frames.

   What we did

   We recorded the office cameras, labelled the footage and approved every clip by hand, 2,126 frames in all, with weapon-free stretches so it also learns what is not a weapon. Each person in view now gets a second, zoomed look, so small and distant weapons are checked close up.

   NowA rifle is found in 61 of 84 test frames, a knife in 195 of 197.

2. False alarms

   Things that look like weapons.

   The problem

   A statue behind the desk was boxed as a handgun. A false alarm that keeps coming back teaches a team to ignore alerts, and then the real one is ignored too.

   What we did

   A weapon alerts only when it is held by, or close to, a person, and a person is never taken for a weapon. Firearms and blades each have their own threshold, so quietening a false knife never hides a real gun. Alerts a reviewer marks False alarm go back into training.

   Now7 false alarms in 54 reviewed alerts (14–21 October, before the last of these changes).

3. Testing

   A score that flattered the model.

   The problem

   On 22 October we found that about half of the general test pictures had near-copies among the training pictures. The score was partly measuring memory.

   What we did

   We removed them, and the score fell from 0.82 to 0.70 overnight. Every new training picture is now checked against the test set, and a new model replaces the running one only when it scores measurably better on pictures it has never seen.

   Now20 overnight training runs since 23 October: 3 promoted, the other 17 changed nothing.

4. Cameras

   Real cameras misbehave.

   The problem

   The picture drifted further behind the longer a camera ran. Frames arrived torn or with garbled colour. A still scene looked like a frozen feed, and a camera restarting or going offline could hold up the whole server.

   What we did

   It now reads only each camera's newest picture, so a delay cannot build up. Video comes over TCP, and damaged frames are dropped before the AI sees them. It tells a quiet scene from a stalled feed, and each camera's trouble stays with that camera.

   NowThe live view stays current, and one failing camera no longer stalls the others.

5. Hardware

   One computer, not a data centre.

   The problem

   Everything runs on one machine at the site; in our trial, a laptop. Each look at a camera with a person in view takes about 51 milliseconds of graphics-card time. Overnight training drove the laptop's card to 85–88 °C within half a minute, and twice a night's training ended in a fault.

   What we did

   We measured how many cameras one computer really handles, and published that instead of a guess. Then we made one pass on the graphics card check several cameras' pictures at once, instead of one camera at a time. Training now checks the card's temperature every two seconds, and how fast it is rising, and rests between batches before it overheats. A fault resumes the night instead of ending it.

   NowOn the trial laptop, 1.6 looks a second per camera with ten cameras and 0.8 with twenty-five, six to nine times what it managed one camera at a time, and a recommended computer for each plan.

6. Reliability

   The restart nobody planned.

   The problem

   During the trial, a Windows update reset the computer's virtual network at 2 a.m. The file behind the system's message queue was cut off mid-write, and the system would not start again until it was repaired by hand.

   What we did

   Each start now checks that file and repairs it if it is damaged; we tested this on a deliberately broken copy. The database is backed up daily, and a restore undoes itself if it fails. Detection, on-screen alerts and review carry on without internet, and a licence keeps working for 30 days offline.

   NowA damaged queue file no longer stops the system starting.

7. Security

   Video of a site is sensitive.

   The problem

   A camera system holds video of the premises, faces and number plates. One open door would matter more than a missed alert.

   What we did

   Security reviews in June and October checked who can reach every part of the system. Anything not meant to be public needs a sign-in. Each install makes its own passwords and keys, licences are signed so they cannot be forged, internal services listen only inside the computer, and the video relay admits only the system, its operators and registered drones.

   NowTwo ports are open to the site's network: the dashboard and the drone feed.

8. Installation

   A site with no IT team.

   The problem

   The installer has to set up Docker, graphics-card support, a database and a 4 GB AI engine on a computer we have never seen. Our first clean Linux test found that a supporting service had been withdrawn from the public download registries, and that the package's file permissions broke graphics-card support. On Windows, the built-in PowerShell could not read our setup scripts.

   What we did

   We dropped that service, fixed the permissions, and made the release build check every setup script before it ships. Then we installed the release again on a fresh machine.

   NowOn a freshly made Ubuntu 24.04 machine it installed, ran on the GPU and started its trial first time, in 11 minutes after the download.

What it still gets wrong

• Rifles, about one frame in four

  On the dual-lens camera, a rifle in view is still missed in 23 of 84 test frames.

• Long guns called knives

  Seven of the eight wrong-type alerts were rifles raised as knives. The alert still fired; the label was wrong.

• Ordinary things

  The seven false alarms were a T-shirt (twice), a hand, a necklace, a person taken for a weapon (twice), and one frame with nothing there at all.