Autonomous PDF Invoice Processing

AUTONOMOUS PDF INVOICE PROCESSING

A grounded invoice workflow that generates and reads PDFs, extracts structured fields, verifies every value against document text, reconciles line items and totals, detects duplicates, and posts or quarantines automatically.

MOCK DATA NOTICE Every invoice, vendor, line item, amount, and benchmark label is synthetic demonstration data.

PROOF OF WORK • 10 PDFs generated and processed in the deterministic benchmark • 80 scalar fields compared • 100% fixture-field accuracy and financial reconciliation • 2.31 ms average local extraction time per PDF • 7 automated tests passing

AUTONOMOUS AI PROOF The live model achieved 116/200 strict field-for-field passes. Grounding and normalization accepted 164 cases; 36 remaining cases were repaired automatically. The complete system financially reconciled and machine-approved 200/200 invoices, including 100/100 reordered and decoy-document stress cases, with 0 human approvals.

The raw score is intentionally retained: accounting automation must not trust model output directly. Every publication requires exact document grounding, duplicate protection, and independent arithmetic reconciliation.

HOW IT WORKS PDF → text extraction → AI or deterministic field proposal → document grounding → quantity, subtotal, tax, and total reconciliation → post or quarantine.

REPRODUCE THE EVIDENCE The repository contains generated PDFs, disclosed truth data, all 200 case decisions, machine-readable benchmarks, seven tests, and CI checks.

Repository: https://github.com/Milo318/invoice-data-extraction-poc