TypeSafe AI launches Jev decision model for high-speed classification

TypeSafe AI, a San Francisco-based startup founded by former OpenAI researcher Diogo Almeida, has released Jev, a new category of AI known as a ‘System One’ decision model. Unlike traditional large language models (LLMs) designed for generative tasks like writing or multi-step reasoning, Jev is purpose-built for high-volume, structured decision-making.

Jev functions by taking unstructured input and returning typed, probabilistic decisions. It utilizes three specific question types: Choice (selecting one option from a set), Score (placing input on an ordered scale), and Noul (a yes-or-no response). Because the model provides a full probability distribution for its answers, developers can use confidence scores to automate actions or escalate uncertain cases to humans.

Key advantages of the Jev model include significantly lower latency—reported between 70–500 ms—and lower costs compared to full generative decodes. By providing guaranteed structured output, it eliminates the need for complex JSON parsing or text cleaning. The model is designed for ‘hot path’ applications such as ticket routing, intent detection, and urgency scoring, where speed and reliability are critical.