A weekend proof of concept. Pigeonhole splits classification in two: a reasoning model compiles your description into a precise spec (questions, options, definitions, disambiguation rules and test cases), and a decision model evaluates every request against that spec, returning a typed answer with a calibrated probability for every option.
Runs as two containers, the app and Postgres, with make up. Add a third container to classify locally with Laya instead of OpenRouter.
The support-triage pipeline in the builder: four questions in one model call, a local rule for priority, and the output fields they feed
Try it with a Hindi ticket: Laya picks returns and wrong_item, but with low confidence, so the ticket is flagged for human review