Reading a document is not the difficult part of extraction. The difficult part is establishing when a reading is wrong — reliably, and without a person re-checking every row. That verification problem is where our engineering effort goes.
Extraction fails in specific, repeatable ways. Each has a corresponding engineering control.
A receipt photographed on site and the supplier's emailed PDF invoice describe one expense, yet share nothing as files — different names, formats and dates. Both are entered, and the folder double-counts.
Control — match documents on their content — supplier, amount, date, line structure — never on file identity.
A decimal point on a creased receipt is a few pixels. One engine reads tens of dollars where another reads single dollars, and neither reports any uncertainty. The error is invisible in the output.
Control — run two independent engines and escalate every disagreement for human review.
Payment confirmations, quotes, delivery notes and closing statements are filed alongside genuine invoices and carry figures that look like costs — some of them orders of magnitude larger than any invoice present.
Control — classify document type before any figure enters a total.
A field marked as a total may be pre-discount; a figure beside a tax label may be a registration number. Extraction that trusts labels produces plausible, incorrect data.
Control — validate the arithmetic; where it contradicts the label, the arithmetic governs and the document is flagged.
Documents are read positionally — by word coordinates on the page rather than the order the PDF happens to store them in — so two-column layouts and wrapping tables do not silently scramble the data.
Extracted figures are then validated arithmetically rather than accepted by label: subtotal plus tax must equal the stated total, line items must sum to the subtotal, and transaction pages must reconcile to the control figures the statement itself prints. Where arithmetic and label disagree, the document is flagged for review.
Structured data ready to load into your systems, together with an exception report listing every item the process could not resolve with confidence. Exception reporting is what makes the output auditable — and what allows the remainder to be trusted.