Not a content programme. Notes from building an industrial correlation engine — what broke, what the data actually said, and the decisions we'd defend in a room full of operators.
Every industrial AI demo claims root cause. Ours returns a causal confidence of 0.0 and shows the association instead. Here's the argument for why that's the more useful answer on a plant floor — and why a vendor who claims causation from observational data is telling you something they cannot know.
Draft in progressA walk through one correlation end to end: a rush order, four data sources that never meet, and the seventeen points of headroom nobody could see. The interesting part isn't the answer — it's how long the answer had been sitting there.
Draft in progressA single-site detection vendor sees four unrelated alarms across four buildings. What it takes — architecturally — to see one pattern instead, and why correlation across an estate can't be added to a per-camera product later.
Draft in progressPublished figures go stale the moment the work moves on, and nobody notices. We wrote tests whose only job is to check the numbers in our own documents against the code that produces them — if a published count stops matching what shipped, the build fails.
Draft in progressWe'd rather publish four things worth reading than forty for a search engine. If there's something you want us to write up, say so — it moves up the list.