Invoice & statement extraction

The problem

Manual data entry from thousands of invoices was bottlenecking reconciliation and delaying financial reporting.

The approach

A computer vision pipeline for invoices, bank statements and utility bills, with field-level extraction, validation rules and confidence scoring that routes only genuine edge cases to a human reviewer.

Documents arrive in every layout imaginable — scanned, photographed, exported from a dozen different systems. A single template-matching approach was never going to hold, so the pipeline separates layout understanding from field extraction and treats each page as a set of regions to be interpreted rather than a form to be matched.

Every extracted field carries a confidence score. Above threshold, values flow straight through to reconciliation. Below it, the field is queued for human review with the source region highlighted, so a person confirms one number instead of re-keying a page.

That threshold is the product decision that matters. Set it too low and people stop trusting the output; too high and you have rebuilt manual entry with extra steps. Tuning it against real documents, not a test set, is what moved this from a working model to a working system.