TypeSafe AI · Jev · CMS support desk demo

How Jev makes the call

Jev doesn't write text. You send it some state and a set of typed questions, and it sends back probabilities. Your own code compares those numbers to thresholds and picks the action. Pick a sample on the left and follow the three steps.

  1. InputA support ticket, or a tool call your AI agent wants to run
  2. JevAnswers typed questions with probabilities. No generated text.
  3. Your codeChecks each answer against thresholds you control
  4. ActionRoute the ticket, or allow, confirm or block the tool call
One call. All questions run in parallel on the same state. Judgment to Jev, facts in code. Counting and math stay in code. Confidence gates. Low confidence goes to a human.

This page uses the repo's offline mock (src/mock.js): keyword rules that return answers in Jev's exact response shape. Real Jev reads meaning, so judge accuracy with a live key and npm run demo. The decision rules and thresholds are the real ones from src/triage.js, src/gate.js and src/questions.js.

Infographic: Jev by TypeSafe AI. Input (a user request or data) goes to Jev, which evaluates typed questions (noul yes/no, choice, score) and returns typed results with probabilities. Examples: urgent ticket classification, category selection, severity scoring. Shows where Jev fits between an LLM agent and the action, a comparison with traditional LLMs, and use cases such as support ticket routing, content moderation, e-commerce and agent safety checks.
Jev at a glance. Tap the image to open it full size. The JSON in the image is illustrative; the real response shape is shown in step 2 below.
1

What your code sends to Jev

One POST to api.typesafe.ai/v1/systemone with the state and every question.

2

What Jev sends back

One typed answer per question. The black tick on a bar is the threshold your code checks.

3

What your code decides

Rules run top to bottom. The first one that fires wins.

Thresholds. Move one to see the decision change.