Skip to content
Leon Achata
All work
Documents

InvoiceParser AI

Drop an invoice or receipt in almost any format (PDF, phone photo, XML, e-mail) and watch the form fill itself, with the math checked.

The problem

Accounting teams re-type the same fields from every invoice: issuer, line items, taxes, totals, withholdings. Invoices arrive as scans, photos, spreadsheets and e-mail attachments, and every typo costs time later.

How it works

A LangGraph pipeline reads native text when the file has it and renders page images for photos and scans. A vision LLM extracts a strict schema, then deterministic checks validate line totals, subtotal + tax = total, tax rate, withholding and the RUC check digit, and flag what needs a human look. The form fills field by field in a zero-build web interface, ready to edit and export.

Pipeline

  1. 01Load
  2. 02Clean
  3. 03Extract (vision LLM)
  4. 04Validate
  5. 05Review & export

Where it's used

  • Accounting firms and finance teams processing supplier invoices
  • Expense reports from receipt photos
  • Feeding ERPs with Peruvian facturas, boletas and SUNAT XML

Highlights

  • 17 input formats, including HEIC photos, UBL e-invoices and .eml with attachments
  • OCR-free vision: scans and photos auto-rotated and sent to a vision model
  • Math, tax, withholding and RUC checksum validation with a completeness score
  • Excel / JSON export and local history; any OpenAI-compatible model

Stack

LangGraphVision LLMsFastAPIPydanticSQLiteDocker

Want something like this for your company?

I build production versions of these systems on your data and your infrastructure.

Related service: Document processing

Contact me