Our production team's end-of-day paperwork files itself now

Every shutter ships with paperwork somebody has to file

Every custom shutter we build travels on a paper cut sheet, and our production team marks it up by hand all day - a substituted part here, an inch of flare there, a swapped axle type. Those marks matter long after the shutter ships, so at the end of every day someone sat down and retyped every one of them into three different systems. It replaced the end-of-day steps of three different people, and it was nobody's idea of a good use of skilled time.

Our foreman's time is too valuable for data entry

The person who owned that end-of-day close was our foreman, and his responsibilities kept growing. Every hour he spent keying handwriting into a computer was an hour he was not on the floor doing the work only he can do. The manual data entry had become a tax on our most valuable time. We did not want to just make the typing faster. We wanted it gone.

Nobody becomes a shop foreman because they want to sit at a keyboard.

One scan in, everything else automatic

So we built the whole thing to begin and end with one scan. The team drops the day's stack on the office scanner, and from that single scan the system files the PDF, attaches it to the order in our system of record, writes a per-shutter report row, emails the team a daily proof, and closes out the order the way the foreman used to by hand. It runs fully on its own - an always-on process that checks for new scans about once a minute in the afternoon window - so nobody has to remember to start it.

The goal was to make the paperwork disappear, not just speed up the typing.

The AI reads the handwriting the scanner's OCR couldn't

The hard part was the handwriting. We assumed we would read the scanner's own text layer, but on real sheets the built-in OCR caught almost none of the handwritten marks and garbled a lot of the printed text too. So we stopped fighting it. Instead of reading a bad transcript, we let an AI vision model read the raw picture of the page - the printed fields and the handwriting together - and pull out clean, structured data. That one decision is what made the whole thing possible.

The printer's OCR caught almost none of the handwriting. So we stopped fighting it and let the model read the raw picture.

A working prototype in under two hours

The first working version was fast - a prototype that read a real sheet in under two hours. That speed is the fun part of building with AI, and it is also the trap: a demo that works once is not a system you can trust with your operation. It was live within days, but the real work came after - months of hardening it against the messy reality of a decades-old paper process, including teaching it to heal itself when something it depends on goes down and finish the job on its own once things recover.

The honest part - the AI couldn't tell which way was up

Not everything was clever. For a while, about half the pages came in upside-down - the vision model could read a page beautifully but could not reliably tell which way was up. The fix was not more AI. It was a boring, old-fashioned orientation tool bolted on in front of the smart one, just to point every page the right way first. A real system is usually your fancy new tool plus a few humble ones doing the unglamorous work.

Half the pages came in upside-down. The fix for our fancy vision model was a boring old tool that just points the page the right way.

Design the paperwork away, don't just speed up the typing

Here is the lesson we would give any owner: when a paper process is eating your best people's time, do not just automate the typing - design the paperwork away. Start and end with the thing they already do (in our case, a scan), let the AI do the reading and the matching, and write the results back into the systems you already trust. Built that way, the paperwork does not get faster. It disappears.