A support request arrives with a billing complaint and a looming deadline. Before anyone writes a reply, the application needs to decide which team should receive it and how urgently it needs attention. Jev is designed to help software make those decisions.
What is Jev?
Jev is TypeSafe AI’s first System One model, introduced in early access on September 15, 2026. It is designed to make structured judgments inside software. Give it information about a situation, define the questions you want answered, and receive values that your application can use directly.
Unlike a conversational model, Jev does not compose a free-form reply. It chooses from options you provide, judges whether something is true, or assigns a score using criteria you define. For a customer message, that could mean choosing a department and rating its urgency. Your code then decides what happens next. TypeSafe calls the model family “System One,” with a focus on answering specific questions quickly. Jev can help software choose its next step, while a generative model handles writing or more open-ended tasks.
How Jev turns context into decisions
A Jev request combines the information to examine, called the state, with questions about it. The TypeSafe documentation defines three question types. Choice picks from the options you supply; Score rates the input using your criteria. Both return probability distributions and confidence alongside the answer. Noul estimates the probability that a statement is true, as a value between zero and one.
For the support request, you could ask which department should handle it, whether the customer mentions a deadline, and how badly the problem affects their work. Each question has a narrower job than “handle this customer.” Jev can evaluate several questions independently against the same state in one request.
Separating the questions also makes the workflow easier to change. Suppose billing tickets with an explicit deadline should enter a priority queue. That routing rule belongs in application code. If the policy changes, you can adjust the rule without asking the model to reinterpret the whole support process.
A message that mixes an account problem with a payment failure may not fit one department cleanly. The team could send uncertain cases to a general support queue, choosing the confidence threshold by reviewing representative tickets. That threshold needs testing: even a confident department choice can be wrong.
Where developers are putting Jev to work
Agent workflows offer another clear use case. A browser agent repeatedly chooses an action from the controls currently available. In Jev Ultrafast, the software turns the available page controls into a numbered list. Jev selects an operation and target, while a separate language model produces text when typing is required. The project also checks the selected target before execution. A completed action still needs a separate check that the browser reached the intended result.
Model routing applies the same idea earlier in a request. The community project Jev Codex Router uses Jev to classify a coding turn, then applies a policy to choose a model and reasoning effort. Before adopting this approach, a team needs to check whether the cheaper model still does the job well—and whether retries or switching to another model eat up the savings.
For more projects to explore, browse SeeAPI’s Awesome Jev Use Cases, our independent collection of community projects. You can follow a routing example into its source or explore a different use, such as semantic search or message moderation. Each entry explains what Jev does and what the author’s code or tests actually show. The collection includes experiments as well as integrations, with links to help you judge whether an approach fits your project.
Adding image screening to a moderation workflow
Moderation extends the same idea to a different kind of input. On a website that accepts image uploads, screening the image and deciding which review queue should receive it are separate jobs. The first requires a detector that examines the image itself.
SeeAPI’s NSFW Filter image moderation API provides that image-screening step. You submit a single image URL, then retrieve the result once processing finishes. The completed result includes a flagged value and category labels. Task completion means the check finished; it does not mean the image passed moderation.
In a custom workflow, the detector could screen the image while Jev assesses its caption or a user’s report against a supplied moderation policy, helping route the case for review. Your code would enforce any rules that must always apply and decide what happens next. This is a proposed workflow, not an existing official integration between the two products. If the only requirement is to hold flagged images, ordinary code can handle that branch without another model call.
Choose a first task you can evaluate
TypeSafe bases its speed and cost claims on its own published workflow evaluations. The company says the gains it highlights are near the upper end of what it expects in real use. It also guarantees that answers follow the requested format. That does not make every answer correct: Jev can choose a valid department and still send a ticket to the wrong team.
Start with one decision your team makes often and can check. Run Jev alongside the existing process, save its inputs and answers, and compare the results before letting it trigger actions automatically. Look for confident mistakes as well as uncertain answers, and count the cost of any retries.
For the support inbox, that might mean suggesting a department while a person continues to assign tickets. Once the suggestions are reliable enough, automate the clear cases and keep a person involved in the rest. You can then expand automation based on the results your team has actually seen.