Netra AI runs detection on cameras that were never installed for detection. They were installed years ago, by whoever was cheapest, to look at a gate or a shop floor. Nobody chose them for frame rate, lens, or placement. We work with what is already on the wall.
That constraint produces a specific kind of knowledge, and almost none of it is written down anywhere public. How to get a usable stream out of a rebranded DVR whose web interface hides half its own features. Why a machine that looks stopped on camera is often still running. What happens to alert quality in the second week, once the novelty wears off and people start ignoring their phones.
Two kinds of posts
Some of these are for the person running the factory. What downtime actually costs, where it hides, what you can measure without buying sensors or signing an MES contract. No protocol details.
The rest are for whoever has to make it work: integrators, sysadmins, automation engineers. Camera connectivity, stream formats, detection tuning, notification design. More specific, less polished, with the parts that did not work left in.
What we will not do here
We will not publish numbers we did not measure. Where we cite a market figure or a benchmark we did not produce ourselves, we will say so and you should check it.
We will not publish credentials, customer names, or anything a customer has not cleared in writing. Where a real deployment appears, it appears with permission.
If something here is wrong, tell us and we will fix the post.