The problem

A UK analytics firm needed a document classifier for internal document retrieval and data annotation — but specifically one that was customisable and explainable. A black-box model that classifies accurately most of the time is not useful when an analyst has to justify why a particular document was pulled, or adjust the criteria on the fly for a new client brief.

They needed the rules to be something a person writes and reads, not something a model infers.

What I built

A Django web application for authoring tailored Boolean search rules, which could then be applied at scale across the document corpus. Analysts build the logic themselves — no model retraining loop between having an idea and testing it.

Fuzzy matching, applied selectively where exact-term matching was too brittle — misspellings, transliterations, and variant forms — without giving up the determinism of the rule set everywhere else.

Embedded analytics dashboards scoped to individual rule-sets, so the team could see how a given rule was performing and tune it.

The outcome

  • Increased productivity for the analysts doing retrieval and annotation.
  • Streamlined internal processes, with rule-building moved out of engineering and into the hands of the people who understood the briefs.