Use cases
What Jev is actually good for
Jev replaces the decision steps in your pipeline, not the writing. Each page below is a worked pattern with code, honest caveats, and independent evidence where it exists.
Support ticket triage
MixedAsk one Choice question for the queue and one Score question for severity in the same call — both come back in roughly the time one of them would take, for a fraction of a cent.
LLM model routing
ChoicePut a Choice question in front of your model cascade so a cheap fast model handles the easy majority and only genuinely hard requests reach a frontier model.
✓ Independent evidence
Agent tool selection
ChoiceLet Jev choose the tool and let an LLM fill in the arguments — the selection step stops scaling badly as your tool count grows.
✓ Independent evidence
Content moderation
MixedUse Score for severity tiers and Noul for specific policy checks, then act on the probability rather than a hard label.
Phishing and email security classification
MixedJev's single raw verdict lost on accuracy to a cheap LLM in an independent test — but combining five cheap signal questions from the same call beat it decisively.
✓ Independent evidence
RAG passage classification and reranking
ScoreScore each retrieved passage for actual relevance to the question and drop the ones that only matched on embedding similarity.
Agent output guardrails
NoulAsk several yes/no verification questions about a draft response in one parallel call, and block or regenerate when any probability crosses your line.
Lead scoring and qualification
MixedScore fit and intent against your own rubric, and route the hot ones to a human immediately rather than in tomorrow's batch.
The pattern underneath all of them
Every use case here is the same move: find the step in your pipeline where the answer is a decision rather than a sentence, and stop paying a generative model to make it. Then — because extra questions are nearly free — ask all the adjacent questions you never bothered to ask.
The phishing benchmark makes that second half concrete: one big question underperformed, but five cheap signal questions from the same call, combined in a simple classifier, beat it outright.