Jev Crash Course: Typed AI Decisions with TypeSafe
This Python tutorial uses Noul, Choice, and Score questions to route support tickets. It demonstrates request configuration, confidence gates, composite priority scores, and parallel questions that inform a handoff to application code, an LLM, or a person.
Original by Telusko Alien CoachAgent workflowsIntermediate40 min 17 secPublished
Jev structures output using three primitives: noul (binary probability), choice (categorical pick with probabilities and confidence), and score (probability-weighted position across ordered levels).
Confidence measures probability distribution sharpness (peaked vs. flat) rather than the selected option's probability, enabling explicit code thresholds for autonomous actions versus human review.
Multiple atomic evaluation questions can be fanned out in a single call over identical state in parallel, letting downstream code branch or compute composite normalized priorities.
Worth knowing
High confidence and typed outputs do not guarantee correct real-world decisions; thresholds must be tuned per business risk, and score/choice descriptions must avoid overlapping text.