Eleven Indian banks’ PDF statements, one structured credit picture

The problem

An underwriter reads a PDF bank statement line by line. Indian statements share no format: headers differ, the first page often differs from continuation pages, business names appear with no label. The work is slow and it’s exactly where human attention degrades.

The naive fix is OCR plus table detection, and understanding why it fails is the whole problem. Deciding whether a line is a new transaction, a wrapped continuation, a header or a footer isn’t recognition. It’s structural reasoning under uncertainty, from several weak signals at once: does it match a date pattern, does it carry an amount format, does it contain a transfer keyword, where does it sit horizontally.

What we built

The issuer is identified before parsing begins: a first pass separates account details from the transaction table using keyword density rather than fixed anchor text, because no anchor text is reliable across banks. The IFSC resolves the issuer, which selects the parser.

Extraction runs per bank; results import through a per-bank column mapping held as configuration. Every transaction is then classified into a spend or income channel by tokenising the narration and matching against a keyword table.

The research behind it

Four approaches were explored: table extraction with camelot; computer vision using morphological dilation with kernel sizes derived from measured letter width rather than fixed constants; a TF-IDF and random-forest header classifier alongside a BERT classifier, with a hybrid proposed where the fast classifier handles confident cases and BERT is called only for semantic near-misses; and a fine-tuning trial of a Table Transformer (a table-recognition model) on hand-annotated statement regions.