Machine learning on live plant data, governed like plant.

Why people call

• A model works in a notebook and nobody can say what happens when it is wrong on a Tuesday.
• Inference needs to run near the asset, and the only proposed path opens a new route into the control network.
• A model has been in production for a year and no one has checked whether the plant still resembles its training data.
• The change process for a model is email, while the change process for a PLC is a formal procedure.

What we design

Where inference runs

Edge, above the boundary, or in the cloud — decided by latency and consequence rather than by preference, and placed so it opens no new inbound path.

The return path problem

A model that only reads is an architecture problem you can solve. A model that writes back to the process is a different problem, and it needs the same rigour as any other control change.

Pipelines under change control

Versioned data, versioned models, and a rollback that works — governed to the standard the plant already applies to itself.

Monitoring as a control loop

Drift and data quality watched the way process variables are watched, with a defined response when they move.

What you are left holding

The work survives the consultant leaving, or it was not architecture.

• Model lifecycle design, from data to deployment to retirement
• Placement decision with the latency and consequence reasoning recorded
• Return-path position, if the model writes anything back
• Monitoring and drift-response plan
• Change-control alignment with existing plant procedures