Digitizing 1,000 Manufacturing Drawings & Tickets a Week — A 4-Week Build Log

The Paper Stacks Next Door

Last episode covered an automated ad-video generation workflow — a scene in which the marketing team's desks finally got lighter. Yet if you walk two doors down at the same factory, into the production control office, the picture changes completely. Printed PDF drawings and hastily scrawled work tickets sit in mountains on every desk.

A 50-person parts manufacturer based in the Hwaseong industrial belt in Gyeonggi Province. Machining and assembly are the core business, and roughly 1,000 drawings, purchase orders, and work tickets land on the office desks every week. A single data-entry employee transcribes all of them into the ERP full-time. The plant manager put it this way: "If she quits, we shut down again."

This piece is a build log of how we digitized those 1,000 sheets over four weeks using the same approach as the video workflow in the previous episode — that is, without grinding people down. We've recorded the quantitative numbers and the decision-making details together, so they can be lifted straight into an internal approval deck.

Choosing the Stack — Claude Vision + n8n + Supabase

At first, we assumed Claude Vision alone would handle everything. That assumption broke within four days.

According to figures Anthropic published in April 2026, Claude Opus 4.7's visual recognition accuracy has climbed to 98.5% — nearly a four-fold jump from the 54.5% of the previous version (4.5). Handwriting word error rate (WER) is also reported at around 4.2%. From a columnist's vantage point, the numbers are undeniably seductive. The view from the shop floor, however, was a bit more complicated. On purchase orders that mixed structured tables with stamps and seals, dedicated OCR tools like Tesseract or Document AI still held a stable edge. So we let go of the ambition to run every document through a single model.

The final stack was straightforward. n8n classified and routed documents by type, the Claude Vision API handled body extraction, and Supabase stored the results and indexed them in a searchable form. n8n's retry node played an unexpectedly large role during the late-night moments when the API would briefly wobble. At 2 a.m. on the Wednesday of week four, when the Japan-side API response was delayed for five minutes, we were asleep — but n8n was already reprocessing the queue automatically.

The Numbers at Week Four

Here are the quantitative results after four weeks.

Extraction accuracy on printed drawings landed at 99.1%. Work tickets mixed with handwriting came in at 94.3%. In week one, the handwritten tickets were stalled in the 88% range. It was only after we introduced a separate human-review queue in week four and added eight pages of frequently appearing abbreviations and unit notations as few-shot examples in the prompt that the figure recovered to 94.3%. 94.3% sounds modest, but it means 57 out of every 1,000 sheets still need a human's eyes. Accuracy is always most expensive in the last 1%.

Processing speed averaged 4.2 seconds per sheet — about 2,400 sheets per day in theory. Costs ran roughly KRW 38,000 in API spend per 1,000 sheets. At the client's actual volume of 4,000 sheets per month, that comes out to around KRW 150,000. In human-time terms, one full-time data-entry employee was reduced to 0.3 of a reviewer. The remaining 0.7 was reassigned to quality-data analysis work.

Similar second-order effects show up in other Korean manufacturers. One company converted 100,000 quality inspection reports into a digital database in six weeks, cutting historical-data search from hours down to seconds. What came after was the more interesting part: the digitized data was recycled as training data for an AI quality-prediction model. OCR is not the finish line, it is the starting line.

A Memo for the Decision-Makers Who Have to Sign Off

To the executives who greenlit this project, we underlined three things.

First, 100% accuracy is not the goal. A design that attaches a human-review queue at the 94% mark beats squeezing the last point out to 99% on both time-to-deploy and cost. Second, do not entrust every document to a single model. With a routing layer in place, swapping models or onboarding new document types becomes possible without downtime. Third, design for the second-order use of the digitized data from day one. Stopping at simple ERP-entry replacement leaves the ROI thin. The moment you fold search, prediction, and quality analysis into the picture, the numbers on the approval memo change.

According to data from Korea's Ministry of SMEs and Startups, smart-factory adoption among domestic SMEs sits at 18.6%, and at just 8.7% for micro-businesses. Put differently: more than 80% of Korean manufacturing still runs on paper. To an executive sponsor, that number is both a threat and an opportunity.

What's Next

The next episode will cover a full-stack build log of shipping the same client's internal system on Next.js, Supabase, and Vercel in 28 days. It's the story of how, once the drawings were digitized, the data flowed all the way to the screens where users actually search and reference it during the approval process.