Ideas

Most triage bots pick a queue and hope. A calibrated model can decline to answer.

Ask three questions in one call: which team should own this ticket (Choice), how frustrated the customer sounds (Score), and whether it is a repeat contact (Noul). Then route on confidence rather than on the label alone. High confidence goes straight to the queue. Medium confidence goes to the queue but flags for a human glance. Low confidence goes to a person.

The interesting part is that the threshold is a product decision, not a modelling one. Refunds can demand more confidence than password resets, and you tune that in your own code without touching the model.

Why it suits Jev: every question is a judgment, the answers arrive together in one round trip, and the confidence value is the actual product feature.

The usual approach to record matching is a similarity number and a threshold you tune forever. There is a better shape.

Write one Score whose levels are the decisions: "clearly the same record, merge them", "related but distinct, leave them linked", "different records". Now there is no threshold to fit, because each level already names what your code should do. The answer maps onto a branch.

This generalises well beyond deduplication — any time you find yourself converting a score into actions with cutoffs, consider making the actions the levels instead.

Why it suits Jev: it turns a calibration problem into a modelling one, and the returned probabilities tell you how close the call was.

Run a Score over a diff and its description: how risky is this change, from "comment or formatting only" through "touches auth or payments". Add a Noul for whether the change includes tests, and another for whether it modifies a migration.

Post the result as a check. Low scores merge on one approval; high scores demand two and a named reviewer. Because the levels are ordered and described, the threshold lives in your CI config where anyone can read and argue with it.

Keep the arithmetic out of the model. Count the changed lines yourself and let Jev judge only what a human reviewer would judge.

Why it suits Jev: it runs on every push, so cost and latency matter, and the output feeds a gate rather than a person reading prose.