Bank & Broker PDF Parser

Bank & Broker PDF Parser is a Dilijent Systems applied-AI engineering project for bank and broker statement processing, where every figure has to reconcile exactly.

The challenge

Bank and broker statements arrive as PDFs in a different layout from every institution, and often as scans. The numbers have to be exact, because a single wrong amount breaks the reconciliation an accountant or auditor depends on. Asking a model to transcribe a statement can produce a value that looks right and is not, which is the worst possible failure mode in finance.

Our solution

• Claude Sonnet maps the column layout (date, description, debit, credit, balance) and classifies spending categories, and nothing more
• No RAG: text is extracted directly from the PDF with pdfjs, with an automatic OCR fallback for scanned statements
• Deterministic code extracts every row and mathematically reconciles the running balance, so the model never produces an amount
• Balance maths is verified during extraction, so a mismatch surfaces immediately instead of flowing into a report
• Low token cost by design, because the model maps structure rather than transcribing the document

The result