Turning a hundred purchase orders a day into one import file

Context

My client received purchase orders from counterparties across India, each state issuing them in its own layout, many as scans rather than structured files. A person spent six to eight hours every day reading them and retyping the contents into the format their own software could import.

Leadership & decisions

One schema, not many parsers. The obvious approach is a parser per format. It is also a maintenance trap: every new counterparty layout adds permanent work. I directed the team to normalise everything into a single unified schema, so a new format is a model problem rather than an engineering project.

Vision first, language second. Scanned documents defeat text extraction. Reading the page with a vision model and then normalising the result with a language model handled the scans that a text-only pipeline could not.

Designed for review, not for autonomy. I was explicit with the client that this reduces manual work rather than eliminating it. The system surfaces confidence and defers to a human on anything uncertain — the target was a fast review, not an unattended pipeline. Promising full automation would have been an easier sale and a worse system.

Outcome

Around a hundred orders a day flow through the pipeline into a single import file. A task that took one person six to eight hours a day became a thirty-to-sixty-minute review.

The same architecture was then deployed for a national radio broadcaster handling order extraction at lower volume, with a comparable result.