Turning a bank statement PDF into a spreadsheet with a vision model

Bank statement PDFs are deceptively hard to parse. You'd think it's a
table, but:

• A single transaction "description" frequently wraps across several lines, so you can't just split on newlines.
• Debits and credits flip sign depending on the bank.
• Scanned statements have no text layer at all.

So I didn't try to parse the PDF structure directly. Instead:

1. Render each page to an image with mupdf compiled to WASM.
2. Send the image to a vision model with a strict JSON schema:

{"bankName": ..., "accountNumber": ..., "summary": [...],
"transactions": [{"date": "...", "description": "...",
"amount": 0}]}

1. Ask for negative amounts on debits and positive on credits.
2. Stream rows back over SSE so the UI fills in live.
3. Reassemble and export (xlsx, CSV, and plain-text OFX/QBO/QIF for QuickBooks and Xero).

The things that actually moved the needle on accuracy:

• Separate prompts: a "header + summary + transactions" prompt for page one, a transactions-only prompt for later pages.
• Telling the model "return only JSON, no markdown".
• Explicit sign conventions instead of relying on the model's guess.
• Generating a deterministic FITID for QBO/OFC so re-imports don't duplicate entries.

It's live and free at statementexcelconverter.com. The parts that
still need work: hand-annotated adjustments (paper statements with
notes), and non-Latin statement layouts.

Happy to share more detail or the prompt structure if useful.