TypeSafe’s Jev is an AI for building text judgments, classifications, and scores into apps. Here’s a look at Jev.
Jev is an app’s decision-maker, not a conversation partner
The quickest way to understand Jev is to think about what you want an AI to return. Rather than “Explain this text,” the questions are “Is this text an ad?”, “Is it work-related or personal?”, and “How relevant is it as a search result?” Jev handles questions like these, where the answer format can be defined in advance.
Conversational AI is generally known for answering questions in text. Jev, by contrast, is designed for developers to use inside software. Rather than having users talk directly with Jev, it is meant to sort information and route processes behind the scenes in the apps they use every day.
The point of connection is an API. An API is an interface that lets an app request a task from another service and receive the result. An app sends Jev text and a question, receives a decision, and uses it to determine what to do next.
It returns decisions, choices, or scores
The materials describe three main uses: making a judgment, such as deciding whether something is an ad; choosing from options, such as folders where something could be saved; and assigning a score, such as the relevance of a search result. Because these AIs make decisions, they are described as “decision models.”
For example, an email-organizing app could take the message body and choose a category from “Work,” “Personal,” and “Other.” This is an example to explain the approach: instead of reading a long explanation and extracting the category from it, the app receives the answer it can use directly.
Standard large language models can also perform classifications and produce output in a specified format. The difference is not whether they can classify, but that these models are designed and offered primarily for making such decisions. However, the materials provided do not establish how Jev is built internally or how its performance compares with general-purpose language models.
Reducing small hassles in search and video viewing
Examples of development projects include a browser extension that skips sponsored segments in YouTube videos, a feature that removes negative comments from chats, Gmail search tailored to users’ intent, and task filtering. The materials also describe attempts to change how users interact with copy and paste, drag and drop, and in-page search.
What these uses have in common is selecting information that already exists, rather than creating new text. Keep the parts you want to see, find the email you’re looking for, or narrow down the items that need attention. This suggests a shift in AI’s role, from chat windows to everyday interactions.
However, the development examples described are not the same as stable product features. For example, the results of removing “negative comments” depend on how the system distinguishes harassment from legitimate criticism. Before adopting such a feature, it is necessary to decide what should be excluded and check the results, including cases where content was removed by mistake.
AI’s job is not limited to generating text
Materials dated October 10, 2026, describe several companies offering decision models in a format similar to Jev. These return answers that software can readily use as-is, such as the probability that a condition is met, a selection from a set of options, or a rating on a scale.
The backdrop is that AI does not need long text at every step when carrying out multiple tasks. At decision points—such as choosing which process to use next or checking whether a result meets a condition—a selection or judgment value may be enough to keep things moving.
Jev’s approach is to put decision-making functions inside apps rather than treat AI as a conversation partner. But a simple answer format does not necessarily mean the decision itself is simple. The same materials also cite early assessments that challenges remain with the consistency of decision models and complex decision-making.
When adopting it, define the boundaries of its decisions before focusing on speed
For practical use, it makes sense to first consider separately what the AI should decide and what happens if it gets something wrong. Misclassifying an email with a label has a different impact from automatically hiding a comment. It is also worth considering a design that sends ambiguous cases to a person.
A short answer does not necessarily reduce the cost of the overall system. The materials also point out that integrating Jev to compress the conversation history of a coding assistant AI could undermine a mechanism that keeps costs down by reusing past inputs.
The materials provided do not include Jev’s specific pricing, independent accuracy tests, or terms for storing and using data. For an adoption involving work emails or customer comments, these terms and the rate of errors on actual inputs need to be checked separately.
Comments