Bank statement extraction — PDF and image formats to structured CSV

The challenge

Bank statements arrive in dozens of layouts — digital PDFs, scanned images, multi-column formats, mixed currencies, and varying date conventions depending on the issuing institution. Manual data entry was slow and error-prone, and off-the-shelf OCR tools failed on low-quality scans and complex table structures. The client needed a reliable pipeline that could handle format variation at volume, flag anomalies, and output clean data without human re-keying.

What we built

We built a document ingestion pipeline that accepts PDFs and image files, detects format type, and routes each document through the appropriate extraction path. For digital PDFs, structured text extraction captures tables and fields directly. For scanned documents, a machine-learning OCR model handles layout detection, table boundary recognition, and text extraction from low-quality scans. A post-processing layer normalises date formats, currency symbols, debit/credit notation, and merchant names across institution styles. Extracted rows are validated against expected totals and flagged for human review when confidence falls below threshold. Output is a clean, consistently structured CSV — with one row per transaction and standardised column headers.

The results

• Statements from multiple institutions and formats processed through a single pipeline
• ML OCR handles scanned and low-quality documents that rule-based tools rejected
• Post-processing normalisation means downstream systems receive consistent data regardless of source format
• Confidence-based flagging keeps a human in the loop only for genuinely ambiguous extractions
• Manual data entry eliminated for the high-volume, high-confidence portion of the document set