What you can build with TypeSafe AI Jev

TypeSafe AI's System One model, Jev, returns typed decisions and probabilities that application code can act on, according to TypeSafe AI. The model supports choice, score, and boolean question types through the AI SDK evaluation interface, enabling use cases such as form routing, ticket prioritization, tool-call review, response-model selection, document categorization, moderation review, and response evaluation. TypeSafe AI advises using Jev for decisions that require interpretation while keeping fixed business rules in code.

Jev, TypeSafe AI's System One model, returns typed decisions and probabilities that application code can act on. You can use it to route forms, prioritize tickets, review tool calls, select response models, categorize documents, flag content for moderation, and evaluate generated answers against supplied requirements.Start with a decision your application already makes, such as assigning a request to a team or flagging a reply for review. Define the possible answers, then test Jev's decisions against examples from your own workflow.Copy link to headingWhich question type fits each use case?AI SDK's evaluation interface exposes choice, score, and boolean questions.Use choice questions to select from defined options and score questions to rate evidence against ordered descriptions. Boolean answers estimate the probability that a statement is true.Use caseEvidence to supplyQuestion typeWhat your application doesForm routingSubmitted fields and destination criteriaChoiceAssign an owner or request another reviewTicket prioritizationCustomer report and relevant service contextScoreSet a suggested priorityTool-call reviewProposed tool name and argumentsChoiceAllow execution or pause for approvalResponse-model selectionRequest context and model descriptionsChoiceInvoke the selected modelDocument categorizationDocument text and category definitionsChoiceAdd a category to the document recordModeration reviewContent, surrounding context, and policy criteriaBooleanAdd a flag for a moderatorResponse evaluationUser request, generated answer, and reference materialBoolean or scoreRecord evaluation results or request revisionUse Jev for decisions that require interpretation, and keep fixed business rules in code. Your application determines what happens next, including whether a selected destination should receive a message.Copy link to heading1. Route incoming forms to the right teamConsider a customer who writes, "I need last month's invoice, but I can't sign in to download it." Sending every message that mentions an invoice to billing would miss the immediate access problem.Ask Jev which destination can resolve the main request. Supply the full form and descriptions that distinguish account access from invoice questions. Include a triage destination for submissions that lack enough information to establish an owner.The Jev and AI SDK Form Router guide implements this pattern for lead, contact, and issue-report forms. It accepts Jev's selected destination when its confidence statistic meets the configured threshold. An uncertain result or evaluation failure sends the submission to a fallback model for an independent decision.For the example above, account access would be the intended first assignment. Include submissions such as "Help with my account" in your test set to check how the application handles missing context.Keep the receiving inboxes in application configuration. The model selects an allowed destination; your code maps that destination to its recipient.Copy link to heading2. Prioritize support tickets by reported impactTicket ownership and ticket priority answer different questions. Two reports can belong to the same team while requiring different response times.Suppose one customer reports a misaligned button and another says every attempt to save a project fails. Ask Jev to assess the reported impact against descriptions such as cosmetic issue, impaired feature with a workaround, or blocked work with no workaround.TypeSafe's Score question supports ordered levels and returns a score that can fall between them, together with probabilities for the levels. Your application can use that assessment to suggest a queue priority. Before automating priority changes, test decision thresholds against the consequences of a mistake.Keep contractual response deadlines in code. Severity assessments interpret the report; they don't establish the customer's service entitlement.An emphatic "URGENT" attached to a cosmetic complaint is a useful test case. Compare it with a calmly worded report of blocked work. The desired behavior depends on the impact rubric, so make those descriptions specific before judging the results.The Jev and AI SDK template demonstrates how to select a team for a support ticket and assess its severity in a single request.Copy link to heading3. Review agent tool calls before executionAn agent's shell tool can inspect or delete a file. The tool name alone doesn't tell you whether a person should approve the operation.In eve's automatic tool-approval workflow, Jev reviews a proposed call and selects either clear or caution. The helper runs clear calls automatically and pauses caution calls for human approval. Failed reviews also require a person.For a policy that permits reading ordinary project files, inspecting a release note could qualify for automatic approval. If the command reads the note and then deletes another file, require review of the deletion even though the first operation is allowed.eve's helper uses the selected choice. Applying a probability threshold requires a custom policy, so don't assume this workflow uses the same acceptance rule as the form router.Within an agent loop, Jev can review proposed actions while application code enforces tool permissions at execution time. The classifier's judgment cannot give the agent additional access..Copy link to heading4. Select a response model for each requestYour help assistant might answer brief product questions and investigate failures spanning multiple services. You can define a model option for each kind of work, then ask Jev to choose between those options before generating a response.eve's automatic model selection defaults to Jev and chooses from an authored allowlist. It evaluates recent text against the model descriptions before inference, retaining the selection through the tool-loop steps in that turn. eve selects again when a new turn begins.For a proposed configuration, describe one option as handling questions answered by a supplied reference and another as handling investigations that require comparing evidence. "What does this setting do?" and "Why does this setting behave differently across environments?" become useful contrasting test inputs.Measure the resulting answers as well as the selections. Choosing the intended model is only one part of success, and adding a routing call doesn't by itself establish a cost or latency improvement. Compare the complete workflow against using your existing model for every request.Copy link to heading5. Categorize documents for search and organizationTo build search filters for a documentation library, you may need to distinguish tutorials from troubleshooting pages. File names alone may not reveal the document's purpose.Jev's choice questions select from categories you define. In a proposed document-classification workflow, supply the document text and describe each category by what the reader can accomplish with it.Pages that explain how to diagnose a failed import could belong under troubleshooting even if they contain numbered instructions. Tutorials walk the reader through completing a task. Decide how to treat documents that serve both purposes, and include an option for material outside the taxonomy.Your application would store the accepted category as metadata. Keep the original text available so editors can inspect disagreements and correct labels without losing the source.Jev currently accepts text-based inputs. Extract readable text from scanned documents before evaluating them, and check the extraction when a classification seems wrong. Missing headings or broken tables can change the evidence the model receives.Copy link to heading6. Flag content for moderation reviewOn a community forum, you could use Jev to flag posts that need attention under a written policy. Start with a narrow question, such as whether a post appears to promote an unrelated service in a support discussion.The underlying mechanism is a yes-or-no evaluation with a returned probability. TypeSafe calls this question type Noul; AI SDK exposes it as boolean.For this proposed workflow, supply the post with enough thread context to distinguish unsolicited promotion from a relevant recommendation. Write criteria that allow a user to answer "Which service worked for you?" without treating every product mention as advertising.The application can send flagged posts to moderators while retaining the question and evidence they need to review the flag. Give reviewers a way to correct the result.Test missed violations and unnecessary flags separately. This example proposes a review queue, with no claim that Jev replaces a moderation system or has demonstrated accuracy on your community's policy.Copy link to heading7. Evaluate generated responses against requirementsSupport replies can sound appropriate while leaving the customer's question unanswered. Evaluate the draft against the original request and any reference material needed to judge it.For example, a customer asks whether changing plans preserves saved projects. Explaining the upgrade process without addressing the saved projects leaves that question unanswered. Ask a boolean question about whether the reply addresses the concern, and use a separate score to assess how clearly it explains the relevant policy.eve's judge integration uses Jev by default to score evaluation assertions. It accepts written criteria or typed questions, including batches evaluated against shared state. This provides a documented starting point for checking responses in an evaluation suite.Begin with replies that reviewers have already assessed. Compare Jev's judgments with those labels and examine disagreements before using similar checks to hold a live reply for revision.The reference you supply determines what evidence the check can use. To assess whether the reply matches an account's actual state, supply the relevant account record or verify the claim directly in code. Agreement with a policy page cannot confirm that a particular upgrade succeeded.Copy link to headingWhat should stay outside Jev's decision?Jev doesn't generate replies or explanations of its reasoning. Keep a generative model for drafting messages, and retrieve authoritative records before evaluating claims that depend on them.Use deterministic checks for exact conditions, such as whether a destination exists or a user has permission to modify a file. Even when the answer fits the allowed type, the model's judgment can be wrong.Choose an initial application where you can inspect and correct outcomes. For form routing, record the assigned owner and later corrections. For response evaluation, retain reviewer labels alongside the model's judgments. Those records let you assess whether the decision is useful in your workflow before expanding its role.Copy link to headingFrequently asked questionsCopy link to headingCan I use Jev in an application without an AI agent?Yes. You can call Jev from a form handler or document-processing service without running an agent loop. The surrounding application still supplies the input and implements the action associated with the answer.Copy link to headingDoes document categorization require Jev to create new labels?No. Choice questions select among categories you provide. If your library needs a different taxonomy, decide on those categories before using Jev to assign documents to them.Copy link to headingCan Jev grade responses from another AI model?Yes. You can supply a generated response with evaluation criteria and relevant references. eve's judge integration supports this in evaluation suites, but the resulting grades still need comparison with human assessments.Copy link to headingWhich Jev use case should I try first?Start with a decision for which you already have examples and a way to correct mistakes. Routing forms into a reviewable queue is a practical candidate; the published form-router template provides an implementation to adapt.Copy link to headingNext stepsFollow How to classify, route, and score with Jev and AI SDK to add typed evaluation questions to your application.Build a form-routing workflow with Jev and AI SDK, including a fallback model for uncertain decisions.Configure automatic tool approvals in eve to review proposed actions before they execute.Compare when to use Jev, deterministic code, or a chat model when choosing which decisions to automate.Plan where Jev fits in an agent loop, with tool execution and permissions controlled by application code.Choose decision thresholds by testing Jev’s probabilities against labeled examples and the consequences of mistakes.Add response evaluations with eve’s judge to assess generated answers against your criteria. Interpret Jev's probabilities, confidence, and scores, then test decision thresholds against the cost of mistakes in your application. Place Jev in an AI agent loop for bounded decisions, while application code controls tool permissions, execution, and state updates.