A sensor on the production line starts drifting. Somewhere inside hundreds of terabytes of logs and telemetry the answer is already there. So why is it still so hard to get it out?

This session goes under the hood of a real industrial predictive maintenance system on Google Cloud. Sensors streaming off the line, the ingestion and processing pipeline, models on BigQuery ML and Vertex AI, and the operational layer where an engineer finally asks a question and gets an answer back.

Then we reach the interesting part! We ground the model on real maintenance documentation so it stops inventing and starts citing its sources, and we watch it diagnose a failure by combining sensor data with the manual. And then I ask the same question badly on purpose, and the model politely agrees with me.

That is the part nobody puts on a slide. In an industrial context the biggest risk is not a model that is wrong, it is a model that wants to please you. On a production line, inside a safety procedure, in a data integrity check, a confident and agreeable answer is a hazard. We'll explore how to design prompts, reasoning steps and grounding so the system is allowed to say "I don't know" and even "you are asking me the wrong question".

We'll close with multimodal, where image, video and audio turn into anomaly detection, PPE checks and presence checks on the shop floor.

You'll leave knowing how an industrial AI pipeline is actually built, and how to stop it from being agreeable exactly when it matters most.

Your production line is drowning in telemetry and the answer is already in there! Under the hood of a real industrial predictive maintenance pipeline on Google Cloud: sensors, BigQuery ML, Vertex AI, grounding on maintenance manuals, multimodal checks on the shop floor. Plus the risk nobody mentions, a model so eager to please that it agrees with a wrong question. Come and see what industrial AI looks like when it runs for real!