Good factory data is messy data.
Clean demo datasets hide the problems agents are built to find.
I went looking for open manufacturing data to build agents on. The cleanest datasets were the least useful.
One synthetic inventory set looked perfect, and it had no disruptions in it. No shortage, no late delivery, no bad week. An agent built to find causes has nothing to find there.
Real plant data looks different. Hours are missing. Operators write “machine issue” and nothing else. One stop has three possible causes. That mess is the job.
So I built my own sample: 14 days of a two-line bottling plant, with real-looking gaps and vague comments, and known events hidden inside.
If you build AI for operations, test it on the data you will actually get.