Demo: Natural Language Operations Assistant
MEASURED RESULTS
THE PROBLEM
When a line stops, the information already exists; it is simply scattered across the MES, Historian, and deviation database. A line manager spends about twenty minutes navigating between three interfaces with three different search methods just to piece together what happened. Meanwhile, the line is down, and some deviations remain open for days because they were not understood quickly enough.
Commercial conversational AI solutions solve access issues but create two deal-breaking problems in regulated environments: production data goes through external services, and there is no way to justify the source of an answer to an auditor.
THE SOLUTION
A natural language query interface connected in read-only mode to all three systems. You ask a question just like you would to a colleague ('Why is line 4 down?') and the answer appears along with its sources on screen: relevant alarms, impacted batches, and linked open deviations.
Two core architectural choices structure the system. The model is containerized and deployed on the site's own infrastructure: no production data leaves the IT system, and no calls are made to external services. Every query feeds a timestamped audit log consultable directly within the interface, turning traceability from a constraint into a demonstration.
The demo runs on a simulated dataset replicating a production environment. It is designed to be re-run on a site's actual data within two weeks.
STACK AND TOOLS