Notes from a hybrid OCR + LLM invoice pipeline

• pdfplumber for digital PDFs; Tesseract OCR (PSM 6, 300 DPI) for scans.
• Gemini 2.5 Flash with strict JSON prompt templates for vendors, invoice numbers, currencies, taxes and line items.
• JSON logs and multi-sheet Excel output with duplicate validation.

Accuracy is not the hard part on eight invoices. Knowing when the extraction is wrong, without a human checking every field, is. That question now drives a lot of my reading in document AI.