Shipping an extraction pipeline that survives real documents

Accuracy on a clean test set is a demo. The interesting work starts at the scanned fax.

TUE 11 AUG 20261 min81 wordswritten at 01:20, KLLOCKED IN☕ × 4

↑ the rig

verdict

Worth it, once you accept that the review queue is a feature and not a failure.

would use again

YES ✓

The demo set was clean scans. Production was a fax of a photocopy of a stamped form, rotated four degrees.

What actually moved the number#

Not the prompt. The confidence threshold, and what happens below it.

if (field.confidence < THRESHOLD) {
  queueForReview(field);   // loudly
} else {
  accept(field);
}

A pipeline that silently accepts a low-confidence field is worse than one that stops, because you find out months later.

↑ was it any good?

↑ pass it on

joe@joeee.dev:~/comments$ tail -f0 entries

reading…