Inticore

Quoting a casting job starts with a customer drawing. An engineer has to read it, list every dimension, spot what is missing, judge whether the part can be cast at all, and write the queries back to the customer. It is slow, and it depends on which engineer picks it up, because the knowledge of how to read a drawing sits in people rather than in a document.

I built a quotation-rules knowledge pack for a custom GPT. When a drawing is uploaded, the assistant tabulates every dimension, specification, tolerance, draft angle and surface finish; flags missing information; checks feasibility against the stated guidelines; lists the extra information needed from the customer; and writes notes for the design team.

The hard part was not the assistant. It was getting the rules out of engineers’ heads and onto a page where they could be argued with. Once written down, the rules are the asset, and the model is just the thing that applies them consistently at 9am and at 9pm.

I also mapped the documents a quote needs and the engineering software already in use, so the assistant fits the existing desk rather than replacing it. And I designed a roughly six-month live-project programme with PSG College of Technology, in which student teams apply machine learning, automation and generative AI on Azure to digitise manual processes at the plant, with an academic mentor and an industry mentor each.

No outcome was formally measured. I do not know how often the assistant’s dimension table was right, how many queries it saved, or whether it shortened a quote. Nobody set up that measurement, which is on me as much as anyone.

The same problem, quoting from a drawing, became an IIM Calcutta teaching case on an aluminium-casting manufacturer, and I later built a fuller agent system around that case as a teaching reference.