85% of messy trade documents handled straight through — measured, not guessed

The problem worth solving

Someone on every trade team retypes supplier documents into the company template — and introduces a typo roughly every hundred fields. The worst offenders are files from Taiwanese and Hong Kong suppliers saved in older Traditional-Chinese encodings: opened with the wrong assumption, the supplier name, addresses, and product lines all come out as garbage characters. The data is intact underneath; it just needs to be read correctly instead of retyped.

The craft decisions

• Exact over clever. Extracted numbers are treated as claims, not facts: every line is recomputed (quantity × unit price), summed, and reconciled against the stated total before anything is produced. A mismatch stops the document — sometimes the supplier's own arithmetic is the error, and catching that before payment pays for the whole exercise.
• Flag, never guess. No pipeline gets every file. The design decision that keeps trust is what happens to the ones it can't handle: they go to a clearly-reasoned exception queue for a quick human check. The 85% that pass go straight through; the rest take a minute of eyes, not a re-do.
• Read the encodings everyone else garbles. The legacy Traditional-Chinese files that usually force manual retyping are read correctly — for many teams that single capability is the difference between "tool" and "we actually use it."
• Print-ready or nothing. Output is a clean, validated PDF or Word file in the team's canonical template, with the original archived beside it for the day an auditor asks.

What shipped

TradeDoc runs in production today, processing real trade paperwork for daily business use. The measured benchmark — about 85% of 186 real documents handled cleanly, the rest flagged or image-only scans — is published on the service page, and the method behind it is written up in our standardization guide. The number is deliberately conservative: we'd rather undersell and be believed.