PDF & Excel — the boring automations that pay

Three PDF flavours, three tools:

1. Text-based PDFs — pdfplumber. Extracts text + tables.
2. Scanned PDFs (images of text) — Tesseract OCR (or paid: AWS Textract, Azure Document Intelligence).
3. Form-based PDFs — pdfminer.six or pypdfform.

Wrong tool = garbage extraction. Always look at the PDF first. Can you select the text in a viewer? It's text-based. Otherwise it's an image.

openpyxl is the standard for reading and writing .xlsx. It preserves formatting, formulas, multi-sheet workbooks.

For pure number-crunching at scale, read with pandas/polars (pl.read_excel), process, write back. Don't loop over openpyxl cells — it's slow.

For templates (write data into a fixed layout), openpyxl wins — you can update specific cells without disturbing others.

pdfplumber's extract_tables returns lists of rows. Drop straight into pandas. The code is 8 lines for a typical financial-report PDF.

Hand-rolled Tesseract OCR is fine for English text. For invoices, receipts, IDs — accuracy drops fast.

Use a service:
- AWS Textract — strong for forms + tables
- Azure Document Intelligence — best for invoices
- Google Document AI — broad doc types
- Veryfi / Klippa — invoice-specific

All have free tiers. The accuracy delta vs hand-rolled is significant; the time saved on QA is huge.

Day 82 done.

- Match PDF tool to PDF type
- openpyxl vs polars for Excel
- 8-line PDF table extraction
- Use a service for invoice OCR

Tomorrow (Day 83): API integrations and the retry/backoff patterns that make scripts reliable.