The warehouse we test everything on

What it was like before

The cameras recorded, and that was all they did. If something went missing, someone lost an afternoon scrubbing through footage, usually without knowing which hour to look at. The recorder's own motion alerts were switched off within a week, because a truck passing on the road at 2am looked exactly like a person in the yard.

What we set up

Seven channels: two warehouse floors, the front, the side, the office, an inside camera and the driveway for number plates. Each camera has its own analysis budget per hour and per day, because the Front camera sees every delivery and would otherwise use up the day's processing before the stock cameras got a look in.

• After-hours sweep. Outside working hours and on closed days, a sweep checks the watched cameras. A person triggers an email with the frame attached, then a cooldown.
• Tripwires and zones across the dock and entries, so crossings are recorded as events rather than buried in footage.
• Clips captured around goods-movement events, about ten seconds before and after.
• Live detection drawn on the dashboard while someone has it open, reading the same local restream the browser is watching.

What we learned

Three things changed how we set the system up for customers.

Budgets matter more than models. Our first version analysed everything it could and used up a day's processing by mid-morning, leaving the afternoon unwatched. Hourly budgets per camera fixed that, with frames near a tripwire crossing always analysed regardless of budget.

Cooldowns are what keep alerts switched on. An alert every thirty seconds is the same as no alerts, because people stop reading them. One email, then quiet for half an hour, is what makes the system survivable.

The office camera is not the warehouse camera. A desk in view means constant movement that nobody needs logged, so that channel looks once every five minutes. Different cameras genuinely need different rules.

What it does for us now

The useful change is small and boring: when something happens, we know that evening instead of next week, and finding the moment takes a minute instead of an afternoon. Goods movements are tied to clips, so a question about a missing item starts with a video rather than a guess.

We do not claim it has stopped anything. It has made the footage usable, which is the honest version of what AI on cameras buys you today.