Invoice Processing & OCR Automation

Clients already emailed their documents, so the pipeline starts there: a monitored mailbox collects attachments automatically, and an upload page accepts files from clients who prefer to drag and drop. Each document is stored and tracked from the moment it arrives.

Background workers built with Celery and Redis process documents as they come in, so there is no daily batch to wait for and the queue never builds up overnight.

• Monitored mailbox and upload page
• Every document stored and tracked on arrival
• Continuous processing with background workers

Where a PDF already contains text, the service reads it directly; where it does not — scans and phone photos — Tesseract OCR converts the image to text first. Extraction then identifies the supplier, invoice number, dates, line items, tax and totals, validated with Pydantic so every field has the expected type and format.

Cross-checks catch common problems: line items that do not add up to the net total, tax that does not match the rate, or a supplier the practice has never seen before.

• OCR only where a document has no text layer
• Supplier, dates, line items, tax and totals extracted
• Arithmetic and supplier cross-checks

Every extracted field carries a confidence score. High-confidence documents post straight into the accounting system, while anything below the threshold goes to a review screen that shows the scan beside the extracted values with the doubtful field highlighted.

A reviewer corrects the value in seconds and approves it. Those corrections feed back into supplier-specific extraction rules, so the next invoice from the same supplier is more likely to post automatically. Accuracy improves with use rather than staying fixed.

• Per-field confidence scores
• Scan and fields side by side for quick correction
• Corrections improve supplier-specific rules

Approved documents post through the accounting system’s API, so there is no export, import or re-keying step. Each posting links back to the original document and its extraction record, including any corrections and who made them.

When an auditor or client asks where a figure came from, the answer is one click away. With 93% of documents posting automatically, staff spend their time on the exceptions and on advisory work instead of data entry.

• Direct posting through the accounting API
• Every posting linked to its source document
• Corrections and approvals recorded for audit