Learn Jev end to end

Most of the work inside an AI app isn’t writing — it’s deciding. Is this email urgent? Is this command safe? Which tool should run next? Jev (by TypeSafe AI) is a model built only for decisions: give it text, a question, and the answers it’s allowed to pick, and it returns one choice with calibrated probabilities. It can’t invent an answer you didn’t list, and it can’t write an essay. That’s the point.

The course teaches the split: Jev makes the many small decisions, the LLM does the few things that need language, and plain code enforces the hard rules. Jev plugs into the agent loop at five places — router, tool, guard, done gate, and judge.

Each notebook runs in about a minute, checks itself against labeled data, and prints what it cost. Along the way you build an email triage job, an SMS scam shield, a code vulnerability hunter, action guardrails for a shell agent, and a model-and-tool router. Notebooks are committed with real outputs, so you can read them before running anything.