Jev AI: Structured Decisions for Your Application
Introduction
Some AI tasks need a paragraph. Others need a decision: which team should receive a support ticket, whether an answer matches its source, or which model should handle a request.
Jev is a model from TypeSafe AI built for that second group. It takes context and questions, then returns structured answers that application code can use. TypeSafe calls this a System One model, referring to fast, focused judgments. TypeSafe’s overview explains the distinction.
Main Content
What does Jev return?
A request contains a state, meaning the information Jev should read, and typed questions, meaning questions with a defined answer format.
There are three question types:
| Type | What it does | Example question |
|---|---|---|
| Choice | Selects an option you define | Which support team should handle this? |
| Score | Rates the input against ordered criteria | How frustrated does the customer appear? |
| Noul | Returns a probability of yes, between 0 and 1 | Does the message express urgency? |
Choice and Score also return probabilities and a confidence value. Noul returns the yes probability without a separate confidence field. You can ask several independent questions about the same state in one call. TypeSafe’s introduction describes these outputs.
Where it fits in an application
Jev supplies a judgment. Your code decides what happens next. For example, a support workflow can use the selected department to route a ticket and send uncertain cases for review. TypeSafe documents this approach in its application patterns.
This separation keeps routing rules in your application. Changing which queue receives a ticket does not require asking the model to manage the whole support process.
Practical use cases
AI evaluation: Check a generated answer against supplied evidence, or judge whether a response follows a policy. Here, Jev acts as an evaluator inside a larger workflow.
Support triage: Classify tickets, detect urgency, and route requests to the appropriate team.
Search and retrieval: Score how relevant a passage is to a question before passing it to a model that writes the answer. This fits retrieval-augmented generation, or RAG, where retrieved information provides context for generation.
Model routing: Classify a request and use application rules to select a model or escalate it for further handling.
These are examples listed in TypeSafe’s use-case guide. They are possible applications, not measured results from a project of mine.
How to get started
Start in the TypeSafe Playground. Paste a support message as the state and add one Noul question about urgency. Inspect the result, then try messages with different wording and ambiguous cases.
Once the question is useful, get an API key from the TypeSafe dashboard. Its quick start documents this request:
POST https://api.typesafe.ai/v1/systemone
Authorization: Bearer <API_KEY>
Content-Type: application/json{
"model": "jev-latest",
"state": "Our integration is failing and we need help today.",
"questions": {
"urgency": {
"type": "noul",
"instructions": "Does this message express urgency?"
}
}
}This is an illustrative request using the documented format, not a recorded test. Read answers.urgency.noul for the returned probability. Keep the key on your server and choose the action threshold using examples from your own workflow.
What to keep in mind
Jev currently accepts text, including text represented as JSON objects or arrays. It does not directly process images, audio, or video, and it does not write explanations or replies. A generative model is still needed when the output should be prose.
A valid answer format also does not guarantee a correct judgment. TypeSafe explains that calibration is measured across groups of predictions; an individual answer can still be wrong. System One documentation
Conclusion
Jev is useful when a task can be expressed as a focused question with a defined answer space. Start with one decision, check it against known examples, and connect the result to application rules. That gives you a practical way to assess whether it belongs in your workflow.