Jev: A Decision-Making AI Layer

Sep 19

Jev is an AI model that provides typed answers with probabilities to questions about a given state, functioning as a 'smart if statement' for automated decision-making.

SignalThree question types: - choice → pick one option (with a probability for each) - score → grade on an ordered scale - noul → yes/no as P(yes), 0 to 1 Choice and score answers also include a confidence value.

Pattern↑ Jev AI decision engine

Jev isn't an LLM. It doesn't write text. It returns decisions. I spent a day hands-on with TypeSafe AI's new model and built 7 runnable examples. Here's what I learned 🧵 (all in one post) WHAT IT IS You send Jev a state (text or JSON) plus a few questions. It answers with typed values and probabilities. Three question types: - choice → pick one option (with a probability for each) - score → grade on an ordered scale - noul → yes/no as P(yes), 0 to 1 Choice and score answers also include a confidence value. WHY IT'S DIFFERENT No paragraph to parse, no JSON to validate. Ask several questions about the same state and they're all answered in one call. Call it a "smart if statement" for judgment calls: which team gets this ticket, is this user about to churn, should the agent retry or ask the user? WHAT I TESTED A free-tier user hits their team-invite limit, visits /pricing 3 times, clicks "Talk to Sales" and never submits the form. Same input, two runs: ✅ Routing → assign_to_sales_rep at 0.78 both times ⚠️ Churn score → 1.15, then 1.31, confidence only ~0.35–0.41 THE TAKEAWAY Confidence is the feature. Automate the answers Jev is sure about. Send the shaky ones to a human. A single churn score wobbling between runs is exactly the case you don't want to auto-act on. COST About 600 input tokens per call. At the documented $0.042 per million input tokens (output is free), that's roughly $0.00003 per decision. WHAT'S IN THE REPO 7 scenarios you can run in a minute: 1. Support ticket triage 2. Agent loop: "what should the agent do next?" 3. Guardrail: check an LLM answer against its source 4. Resume screening 5. Incident triage from logs 6. Fraud risk scoring 7. A deliberately vague message, so you can watch confidence drop Plus a script that repeats a call N times to measure stability, and a simple confidence gate. CAVEATS - Early access, so you need an API key - Probabilities, not explanations. It's a decision layer, not a whole app - I've tested a handful of scenarios, not benchmarks Try it yourself 👇 t.co/Hfb2BfMtOL Building with Jev too? Tell me what you're trying #Jev #TypeSafeAI #AIAgents

Original post