Extracting data from documents

Retyping details from paper into a computer is work nobody wants to do and at the same time nobody can skip. It is slow, dulling, and precisely for that reason mistakes get made in it.

It can be done so that you photograph the document and the system reads the details itself.

Where I know this from

My bachelor’s thesis was about exactly this, the autonomous extraction of data from accounting documents. Commercial solutions charge around four crowns to process a single document. My solution got below six hundredths of a crown per document, roughly seventy times cheaper.

That difference is not because it is worse. It is because a commercial solution has to handle everything in the world, while a solution built for one company only has to handle what that company actually gets in its inbox.

Where it already runs

The same principle underpins the feature in Paving Planner that turns a photo of a hand sketch into a technical drawing. The original idea was simple: let the paver redraw the area in the browser himself. It turned out the redrawing was the bottleneck and most people did not want to get into it. So out of that came a feature that does it for them.

What is worth knowing up front

This will never be a hundred percent. There is always a document photographed at an angle or written by hand so it cannot be read. So it is built with a check, where a person goes over what the system is unsure of, instead of retyping everything.