TypeSafe AI's Jev skips text, scores options in 70ms

Jev is unusual less for its technology than for what it refuses to do. Co-founder and CEO Diogo Almeida worked at OpenAI and is among the authors of the InstructGPT research that laid the groundwork for ChatGPT, and his new startup has stripped away the part of a language model most users interact with: the writing. Developers define the questions and the permitted answers, and Jev returns only a label and a probability. The company says it reaches this by skipping step-by-step text generation and computing several outputs in parallel, which is why adding more questions in the same call barely increases response time.

The speed advantage matters because it changes where such a model can sit. TypeSafe's example workflows describe checks before every response from an AI assistant — did the draft contradict the conversation, did it promise a refund not recorded on the account — as well as sorting requests by topic, spotting buying intent, and judging when a human should take over. In its store example, the software attaches fixed rules to Jev's labels: payment issues go to accounting, refund requests get flagged, unclear cases go to a staffer. Jev supplies the assessment; the surrounding program decides what happens.

The catch is that preset categories and data structures are not unique to Jev — OpenAI offers Structured Outputs for the same purpose — so TypeSafe would need to match competitors on quality, not just speed and cost. Its published evaluations compare four workflows it built itself against responses from other AI models rather than independently verified correct answers, and GPT-6 Astra does not appear among them. Whether Jev's judgments hold up on a given company's tasks looks like the test that decides its usefulness.