Docket is a service desk that runs on the mail it receives. A message arrives, it is read and understood, sorted against a taxonomy of what the business actually deals with, scored for how urgent and how risky it is, and handed to the team that owns that kind of problem. From there it is tracked through to a resolution, with the clock measured against whatever the service agreement promised.
The worked example is a connectivity provider, the sort that sells broadband, mobile and leased lines, because that is a business where the mail is relentless and the difference between a billing query and an outage matters. None of the domain is written into the code outside two files, so aiming it at a different business is an editing job rather than a rewrite.
The part worth dwelling on is what happens when the classifier is not sure. Rather than guessing and routing anyway, uncertain cases go into a review queue and are blocked from any automatic action until a person releases them. A system that quietly guesses is worse than one that says it does not know, because the guess is invisible until it has already gone wrong.
Behind the console sits a FastAPI service with a local language model for summaries and drafts, embeddings for classification where keywords are not enough, and named entity extraction pulling account numbers and circuit identifiers out of the message body. Every state change leaves an audit record, and resolution time is measured from a frozen timestamp rather than from the last row update, which drifts.