Premier Scan

Operations teams had to manually extract and reconcile fields from invoices and hotel folios, then inspect subtle mismatches across names, dates, rates, taxes, and line-item charges.

• Combined OCR and Bedrock vision extraction so scans and native PDFs could be converted into schema-validated records.
• Ran deterministic audit checks for rate, date, name, tax, duplicate-page, and unexpected-charge mismatches instead of asking an LLM to judge every field.
• Built a three-panel review station where operators can inspect confidence, correct fields, approve results, and feed corrections into a learning log.

• PDF or image input is rasterized into page images and checked for duplicate-page signals.
• Vision extraction produces structured candidates validated against a JSON schema.
• A deterministic audit engine checks business rules and stores findings in an audit database.
• The review UI presents the source document, extracted fields, findings, and correction controls together.

• Golden-set regression tests catch extraction and audit regressions before release.
• Confidence values route uncertain fields to human review instead of silent acceptance.
• Synthetic folios and redacted operational data keep public demonstrations confidentiality-safe.

• Deterministic checks add implementation effort but make financial and operational findings explainable.
• The next step is expanding the golden set across more folio layouts and lower-quality scans.

The work progressed from high-volume invoice extraction into a testable folio-audit workflow with measurable accuracy, explicit findings, and operator approval.