Feedback is most useful when a student can still act on it. A writing app might notice that a claim needs evidence while a paragraph is taking shape. A study app might recognize an assignment as soon as a file is uploaded. These experiences depend on lots of small decisions, made quickly enough to keep up with the learner.
That is why Jev caught the attention of the Worldwide Studios education community. Jev is a model from TypeSafe built to make decisions. A builder gives it text and questions, and it returns answers the product can use directly: a category, a score, or a yes or no signal. Its speed and low cost give builders room to try these checks throughout a learning experience.
Estimated cost per output: $0.000021 for a request with 500 input tokens. At that size, $1 would cover roughly 48,000 requests. As of September 30, 2026, OpenRouter lists Jev 1.13 at $0.042 per million input tokens and $0 for output tokens. The total depends on how much text each request contains.
The opportunity is to make feedback available more often: across a folder of course materials, throughout a writing draft, or during a live conversation. On September 28, four AI for Education Fellows and Michael Raspuzzi brought five early experiments to a Worldwide Studios build session. The video above shows what they tried.
Five demos from the community
1. Wally Khan: organize course materials
A student uploads files from several classes. Which are homework, case preparation, readings, or group assignments? Wally tested that question for Grad Buddy using 27 course files. Jev classified 23 correctly with file text and added context, compared with 22 from text alone. An app could suggest a category as each file arrives and ask the student to confirm uncertain cases. His first test also showed errors, including one with high confidence, so those suggestions still need checking.
2. Ashley Hodges: help students see the parts of an argument
Ashley brought a student’s argumentative paragraph into Jev’s playground and asked about its hook, claim, and supporting evidence. Her interest was in helping students understand how an argument works and helping teachers review their own assignment instructions. The preset she tried was designed for viral posts, so the next step is to write questions that match a classroom rubric. This was a playground experiment ahead of any integration into her classroom resource tool.
3. Damian Matheson: connect learning to job skills
Damian asked Jev to check 500 job postings for 15 named skills, then built an interface to explore the results. His larger goal is to help students connect classroom learning with real roles and problems in industry. The demo showed a way to make job data easier to explore. A next test is to check the skill matches and see whether they help students understand how a class project relates to work they might want to do.
4. Aaron Goldstein: make spoken stories easier to explore
Aaron recorded a story and displayed a visual card alongside labels such as myth, poetry, and stream of consciousness. The labels changed across the recording. That opens up questions for education: could a student revisit a story and see how it moves between different kinds of language? Could an educator help a class compare how stories are told? His demo made the classifications visible so people could inspect and improve them.
5. Michael Raspuzzi: make a learning character respond in the moment
Michael showed an animated Plato that reacted as someone typed. Jev helped choose expressions and mouth shapes, and a separate animation system moved the face. The experiment explored how fast decisions could make a character feel more responsive during a learning conversation. It is one example of Jev working inside an education product alongside the systems that handle dialogue and animation.
These were first experiments; the useful next step is to check the results against examples that students and educators trust.
Start with one decision
The TypeSafe Playground lets builders try questions without code. Jev can choose an option from a list, score something on a defined scale, or answer a yes or no question. The official quick start shows how to put those answers into an application.
For an education product, a good first question is one the product already needs to answer. Test it on real examples, check the mistakes, and decide when to ask a learner or educator for help. Confidence scores can help builders design that handoff, but they also need testing.
Which decision could help a learner immediately if an education product could afford to make it on every interaction?