Jev: A Cursory Look and What Makes It Unique

Imagine my workshop gets a hundred messages like this: "My robot won't charge, and I need it for a demonstration tomorrow." I could ask ChatGPT to help troubleshoot the problem and write a thoughtful reply. But before that, our software has a smaller job: send the message to the right team and decide how urgently it needs attention. That smaller job is where a model called Jev gets interesting.

First, ChatGPT is the application we talk to. Language models are part of what powers it. They can generate explanations, write code, and work through problems, with tools adding capabilities such as web search. That flexibility is useful when the answer could take many different forms. I can ask why a robot might stop charging, then follow up with a question about its battery, without defining every possible answer beforehand.

Jev comes from TypeSafe, which calls it a System One model. You give it the situation, called the state, and specific questions with defined answer types. It returns decisions and probabilities that software can use directly. In our workshop, the state could include the customer's message and our routing rules. Jev currently takes text, including structured text. It doesn't inspect a photograph of the broken robot or write the customer a reply.

There are three kinds of question. Choice selects from options we supply, such as repairs, orders, or something else. Score rates something against levels we describe, like low, medium, or high urgency. The third kind returns a probability for a yes-or-no question, such as whether the customer requested a refund. We can ask several independent questions about the same message together. Then our code decides what to do with those answers.

TypeSafe says Jev is trained to make calibrated decisions. Calibration means that, across many comparable predictions, events given an eighty percent probability should happen about eighty percent of the time. That's a pattern we can measure over a dataset, never a promise about one message. Choice and Score also return a confidence value summarizing how concentrated their probabilities are. We can use uncertainty to send ambiguous cases to a person, after testing what works on our own messages.

Language models can classify messages too. OpenAI's API also supports structured outputs, which constrain answers to a defined format. So a tidy output alone doesn't establish that Jev is better. And either approach can produce a perfectly valid answer that's wrong. Our robot message might get routed to orders even though it needs repairs. Defining the possible labels prevents an unexpected label from appearing; it doesn't guarantee the correct label wins.

Jev's appeal is doing lots of focused judgments quickly. TypeSafe reports that it produces probabilities in parallel, and its own evaluations show substantial speed and cost gains in selected workflows. Those results depend on the task, comparison model, and settings. They don't tell us that every chatbot workload will get the same improvement. We'd want to compare accuracy, response time, and total cost on the messages our workshop actually receives.

For our example, I'd try a combined workflow. Jev helps route the message and flag urgency. Ordinary code checks which actions are allowed. A language model helps explain the charging problem and draft a useful response. Unclear cases get another look. The useful choice is where each model fits: open-ended work, repeated decisions, and the handoffs between them. That's how our broken-robot message reaches someone who can help. Subscribe for more ways to make the robot do it.