Turning unstructured submissions into reviewable decisions

The client operates a lending platform where every disbursement request arrives as a pile of documents, photos, and free text. Human inspectors read the pile and make a call. The judgment was good and the throughput was not.

Their production system is a long-lived monolith carrying real money. Modifying it was off the table for a first engagement, and rightly so.

The work also had to be provable before it could be trusted, which meant it had to run against historical submissions where the inspectors' own decisions were already recorded.

A sidecar. It reads a data export rather than the live system, so nothing it does can affect a disbursement. It extracts structured evidence from each submission, classifies the request, and writes the reasoning behind the classification in plain language.

Past inspector decisions became the evaluation set. Every change to the pipeline could be scored against what experienced humans actually decided, rather than against a hand-written rubric.

The architecture kept the production monolith untouched while producing a system the team could evaluate against their own history, which is what made adoption a decision about evidence rather than a leap of faith.

We are not publishing accuracy numbers here. When the engagement produces figures the client is willing to have quoted, they will appear with their source.