The workbench is where analysts profile a domain, test a match model, score suppliers or forecast demand, on master data and transactions together, with the twelve security layers and the audit trail applied.
Ranges, patterns, nulls, candidate keys; proposed column rules.
Label pairs, train, evaluate; features include transaction behaviour, because the transactions are here.
Supplier risk, customer churn, item obsolescence, published as attributes.
Demand per family and plant from the bed's history; results as a version in the pivot.
Move a node, change a weight, re-roll the totals; compare with the live version.
Source against golden, replica against source, invoiced against booked.
A workbench query runs as the analyst, through the same engine as the grid: row filters, column permissions and masks apply, and reads of sensitive columns are logged with a purpose. Publishing a result is a write like any other: rules run, survivorship ranks it, the audit trail records the model version behind it.
Models run inside the estate; row data stays where it is. A frontier model can be called for structure and metadata under the boundary described on the AI-assisted page.
SQL and Python with the common data-science libraries; R through the same runner on request.
Yes, as a job with a version, an owner and a result check, with the same audit as a rule run.
It sits beside them. Deep engineering stays in your lakehouse; the workbench is for analysis that must run under master-data governance and publish back.
Bring two systems that hold the same customers, suppliers or materials. In one working day we connect them, run the match on live rows, review the queue with your steward and publish the first golden view with its API.