Billy

We are a computer vision team. Pose estimation, diffusion models, papers at CVPR with our names on them. So when we first looked at engineering drawings, we thought we had been handed the easiest input in the field. No lighting. No occlusion. No motion blur. Black strokes on white paper, drawn to ISA-5.1, by people who were trying very hard to be unambiguous. Child's play, right? Models are good at photographs because somebody labelled 14 million of them. Nobody has done that for engineering drawings. The largest public P&ID dataset is 500 synthetic sheets, tested on 12 real ones. Not enough to train on, and not enough to tell you whether you are right. So the model learns all of it from almost nothing. A gate valve, a globe valve and a ball valve are the same bow tie. The difference is what sits in the middle: nothing, a filled circle, an open one. Read it wrong and you have told the plant that a throttling valve is an isolation valve. Text printed over a line is not part of the line. A circuit leaving the right edge picks up three sheets later. And the one that cost us the most time: a tag sitting between a valve and a transmitter, belonging to exactly one of them, with nothing in the image to say which. Models that read the world do not read the page. 8,362 drawings and three and a half years in, somebody still walks over with a sheet where a line was drawn in by hand during a turnaround, and asks whether the model should trust it. There is nowhere to look it up. So we go and ask. Last month that was Parth Tulsyan and Vijay, sitting with the process engineers who read these drawings every day. What they brought back went into the model. Nobody publishes that dataset. We build it, one conversation at a time.