· 4 min read
Jev: fast decisions in every call
Jev is a fast decision model from TypeSafe. It does not write sentences; it answers typed questions. Here is where Woise uses it, and why your agent still talks with a language model.
by Woise Team
A voice agent makes two very different kinds of decision. Some are about words: what to say next, how to phrase it, how to answer a question nobody planned for. Others are plain judgements: did the caller just interrupt, was that call a booking, did the agent follow the script.
A language model is good at the first kind and slow and expensive at the second. Jev, a model from TypeSafe, is built for the second. It does not write text. You give it a situation and a typed question, and it answers with a choice and how likely that choice is.
Where Woise uses it
After every call, the post-call analysis runs on Jev. It reads the call and answers the checks you have turned on: the outcome, the built-in checks, and any checks you wrote yourself. No language model is involved in that step.
During a call, Jev judges a barge-in: whether the caller really cut in or just made a small sound. That is a quick yes-or-no, which is exactly what it is for.
Why a probability and not a paragraph
Because Jev returns how likely each answer is, a check can say how sure it is instead of giving a confident-sounding explanation. A check you wrote applies when the model's yes reaches a set level, and you can read that level back on the call.
It also keeps results steady. The same call gets the same kind of answer every time, which makes the reports easier to compare from one week to the next.
What still uses a language model
Everything the agent says. Jev decides; the language model talks. Keeping the two apart is the point: the model that writes your agent's sentences does not have to also answer every yes-or-no question about the call.