Galleries

  • test delete

    What are we imagining?

    Why?

    What might this make possible?

    Make

    What we’re building

    What we’ve tried

    Where it stands

    Repeat

    What we learned

    What we’d change

    What’s next

  • Jev AI: Small Decisions, New Possibilities

    What happened

    Jev AI takes an unusual approach: it gives up free-form text generation and returns answers from predefined options instead.

    Its features page describes structured outputs, confidence probabilities, and response times of 70–500 milliseconds. The company presents it as something developers can call from ordinary code—a way to add judgment to a specific step in an application.

    Why it’s interesting

    At Imagine Make Repeat, we’re interested in the small decisions that help a project work. Which category does this request belong in? Where should it go next? Is there enough information to continue?

    Jev’s approach caught our attention because those questions don’t always need a written response. Sometimes an application needs a selection it can act on, along with an indication of uncertainty.

    There’s a distinction worth keeping clear. Jev describes its predefined outputs as eliminating hallucinations, but choosing from valid options doesn’t guarantee choosing the correct one. We’d want to test its decisions and confidence estimates on real examples before relying on them.

    What could this make possible?

    One idea we’d like to explore is a community resource finder.

    Someone could describe what they need in their own words. A decision model could help sort that request into categories—food, shelter, transportation, or help with paperwork—then route it toward a maintained directory.

    When the system is uncertain, the next step could be a clarifying question or help from a person. The directory would still need current, verified information; classification alone wouldn’t establish availability or eligibility.

    That’s a possible experiment, rather than something we’ve built or tested with Jev. We’d start with a handful of categories and realistic requests, then look closely at where it succeeds, where it hesitates, and where it gets things wrong.

    The question that stays with us: what could we build if small moments of judgment became easier to add to everyday software?

    Check it out at: https://jevai.net/