Article
CommunityJev Integrations & Adapters
Jev Judgments in Akka
Akka's post walks from a Jev triage example - eight questions with probabilities - to an effects API judgment call and an AI Gateway that can route to hosted Jev or a self-hosted OpenJev model.
- Format
- Article
- Reading time
- 3 min
- Published by
- Community
- Author
- Kevin Hoffman
- Last verified
- Oct 2026
- Languages
- Java
- Capabilities
- Requires API key
Highlights
- The effects API sketch sends route as choice, severity as score, and urgent as yes/no in one judgment call whose Judgment class carries the answers.
- TypeSafe's published figures: 70-500 ms end-to-end, 40x-200x faster on System One queries, $0.042 per million input tokens, and free output.
- The workflow table cites 0.114 s and $0.000081 per task against 8.566 s and $0.013880 for LLMs, with TypeSafe calling those the high end of the gains.
- The order example's eight questions return shipping 0.94, repeat contact 0.97, refund threat 0.88, order number 0.99, and high priority 0.83, among others.
- Akka's AI Gateway can route judgment requests to the hosted Jev API or to a self-hosted OpenJev open-weight model in Akka's inference engine.
Watch out
The performance table is TypeSafe's own launch material, which the post labels as vendor evaluations; the zero-hallucination claim is described as following from the output schema rather than being measured, and the LLM error rates come from OpenRouter traffic. Jev cannot extract values such as an order number.
Questions about this entry
Short answers drawn from the linked project or page.
A decision model with its own interface style: instead of streaming text back and forth, you ask questions about some state and get answers with probability scores, a calibrated confidence figure with every answer.
Through the effects API: effects().judgment().state(ticket).question(route, Question.choice(...)).question(severity, Question.score(...)).question(urgent, Question.yesNo(...)).thenReply(). The Judgment class holds the answers and their probabilities, and the API ships in an upcoming SDK release.
Eight questions about a shipping complaint returned answers with probabilities, from shipping at 0.94 and frustrated at 0.81 to order number at 0.99 and high priority at 0.83; Jev could not extract the order number itself, only confirm one was present.
Interrogate the state and message with fast, cheap questions, then use the answers to decide whether a full, slower LLM call is needed, for example extracting an order ID only when Jev is confident one exists.
With Akka Optimize, Jev is available through the gateway, which can route to the hosted Jev API at TypeSafe's portal or to a self-hosted OpenJev open-weight model in Akka's inference engine, choosing a backend by static policy, semantic embedding, or SystemOne.
70-500 ms end-to-end response times, 40x-200x faster System One queries, 0.114 s and $0.000081 per workflow task versus 8.566 s and $0.013880 for LLMs, $0.042 per million input tokens, and free output; the post labels these TypeSafe's own evaluations and demos.
It says the 193.6x and 444.6x workflow figures are on the higher end of real-world gains, the zero-hallucination claim follows from the output schema and was not measured, and the LLM error rates used for comparison come from OpenRouter traffic.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
X: Jev 1.14 is now on OpenRouter
X: Jev lands on the Vercel AI Gateway
X: Jev as an LLM guardrail inside Agentgateway
YouTube: Greg Isenberg and Ryan Vogel on the businesses Jev unlocks, and a path in through the Vercel AI Gateway.
More Jev resources like this
Fly.io's TypeSafeAI connector gives Sprites access to Jev without the API key entering the machine: the Sprites gateway attaches the organization's encrypted key, and each Sprite sets only TYPESAFE_BASE_URL plus a placeholder key.
Jev Integrations & AdaptersArticle#community#fly-io#agents
Community
LiteLLM serves decision models on two routes that share one provider list: /v1/decisions for OpenAI-format bodies and /v1/systemone for System One bodies, with Jev available as the typesafe/jev-latest provider.
Jev Integrations & AdaptersDocs#community#litellm#gateway
Community
A Kafka Connect sink connector that sends each input record to TypeSafe AI's Jev with questions you define, then publishes an Enriched Record holding the original message, Jev's answers and the evidence to trace and deduplicate each evaluation.
Jev Integrations & Adapters#community#github#kafka
Community
smatiolids/kafka-jev-connectorJava
smatiolids/kafka-jev-connectorJavaMastra's classifier primitive defines choice, score, and boolean questions against an evaluation model such as jev-latest, returning a probability for each criteria key so workflow branch logic can act on the answer.
Jev Integrations & AdaptersArticle#community#typescript#mastra
Community
An official Mastra template that checks pricing, changelog, and documentation pages against saved captures, has Jev classify each change with its evidence, and routes it to ignore, record, request review, or alert.
Jev Integrations & AdaptersDocs#community#typescript#mastra
Community
BAML maps function return types onto Jev questions: bool and float become Noul probabilities, enums and literal unions become Choice picks, and one class turns its fields into a single multi-question request.
Jev Integrations & AdaptersArticle#community#baml#dsl
Community
From the Jev guides
Original write-ups on what this entry is for.
From the Jev AI model community
Posts from builders shipping with Jev right now.