Accounts-payable teams spend hours retyping the same fields off PDFs and phone photos of vendor invoices: supplier, invoice number, date, each line item, tax lines, and the grand total. ERPNext gives you a clean Purchase Invoice DocType and strong accounting behind it, but out of the box there is nothing that reads a document. Every invoice is still keyed by hand, transposition errors slip into the ledger, and the month-end backlog grows. The native Data Import tool assumes you already have structured rows in a spreadsheet, so it does not help with the actual bottleneck, which is turning an unstructured image into structured data.

This app closes that gap. We build a proper Frappe application (its own module, DocTypes, and hooks.py) that ingests a scanned PDF or image, runs it through an OCR engine, then passes the recognized text to an LLM extraction step that returns structured JSON: header fields, a line-item table, tax breakdown, and totals. Each captured document lands in a new Invoice Capture DocType with a confidence score per field. From there a whitelisted server method maps the extracted payload into a draft Purchase Invoice (or Sales Invoice), so the ERPNext-native validation, tax templates, and workflow all still apply — we create documents through the ORM, never with raw SQL.

Vendor auto-matching runs against your Supplier master, using tax ID, name, and prior-invoice heuristics to propose the right supplier and flag anything ambiguous for review. An optional email-inbox ingestion path lets AP forward invoices to a dedicated mailbox; a scheduler event (registered in hooks.py scheduler_events) polls the inbox, pulls attachments, and queues them for capture automatically. Low-confidence fields are routed to a human-in-the-loop review queue — a Frappe list view with role-based permissions — where a clerk confirms or corrects the extraction before the draft invoice is submitted. Nothing posts to your books without a human approving it.

Because this is build-to-order, you are not downloading a fixed marketplace app and hoping it fits. We start with a short scoping call, confirm your document types, tax treatment, approval roles, and which OCR/LLM provider you prefer (cloud API or self-hosted), then build your version against ERPNext v15 or v16. You get UAT on a staging site, a rollback plan, and the full Git repository handed over so you own the code. Typical delivery is one working week from confirmed scope, followed by a post-go-live support window.