Creative Jev: jevmeter, jev-paint and a UI Generator

Most of the Jev projects in this series use it for what it was designed for: fast, cheap, typed decisions inside an agent loop — routing, gating, picking the next click. A smaller cluster points the same mechanism somewhere more unusual: creative and generative work. jevmeter scores video transcripts against a rubric and draws the scores as an on-screen meter. jev-paint turns Jev's raw probability distributions into visual paintings. ui-generator-instinct-jev builds UI mockups by having Jev choose from a fixed catalog of shadcn/ui components rather than generate layout code.

None of these projects claims Jev is generating anything in the usual sense — each one is explicit that Jev is choosing or scoring from a bounded set, and something else (video rendering code, a paint algorithm, a component library) does the actual output. That distinction is the throughline worth understanding before you judge what any of them can do.

Key Takeaways

• Jev doesn't generate video, images or code here — it scores or selects. Every project in this cluster keeps generation in a separate, deterministic system.
• jevmeter's scoring is a model judgment, not fact-checking. The README is explicit that sentence scores reflect Jev's rubric-based read, not external verification.
• jev-paint needs a local Python relay because the Jev API "currently disallows browser CORS origins" — a real infrastructure constraint, not a design choice.
• ui-generator-instinct-jev is bounded by design: Jev can only select from an existing shadcn/ui component catalog, so it can't invent new page structures or copy.
• All three are small (7–81 stars) as of September 2026 — genuinely early demonstrations of the pattern, not production tools.

jevmeter: scoring transcript sentences, not judging truth

jevmeter (MIT, 81 stars as of September 2026) creates edited videos with on-screen "meters" by asking Jev to rate transcript sentences against a chosen rubric, then aligning those scores to the corresponding segment on the video timeline. The mechanism is the same typed-question pattern used throughout this series — preset questions and fixed scales rather than free-text — applied to spoken content instead of UI state.

The project's own dataset entry is careful about what this proves: transcript scoring is "a model judgment, not external fact-checking or mind-reading." Author-reported accuracy and cost figures were not independently retested for this article, and video processing being possible doesn't imply the pipeline runs in hard real time — expect batch turnaround, not a live overlay. If you're using jevmeter to grade rhetoric, sentiment, or rubric adherence in a video, treat the meter as one model's read of the transcript, not a verified score. Aiskill.market entry: /skills/jevmeter-chetaslua.

jev-paint: probability distributions as pixels

jev-paint (MIT, 51 stars as of September 2026) is the most literal "creative use" of Jev in this cluster: it takes a text prompt, asks Jev for pixel-level probability distributions, and runs those through an artistic algorithm to produce a painting. It's deliberately minimal — Python 3.9+ standard library on the backend, ES modules and no runtime dependencies on the frontend, requiring only a modern browser with module workers and OffscreenCanvas support (current Chrome, Edge, Firefox or Safari).

One infrastructure detail matters for anyone trying it: because the Jev API "currently disallows browser CORS origins," the frontend can't call Jev directly. A local Python helper (python3 server.py, served at http://127.0.0.1:8791) relays requests instead. Larger canvas grids mean more API calls and more tokens consumed. The project is explicit about being single-user and ephemeral — paintings clear on page reload since there's only in-memory storage, the API key is stored unencrypted in browser localStorage, and the README states outright that this is "not a publicly hosted multi-user service." Rendering happens at 560×560 with a 550ms fade-in. Aiskill.market entry: /skills/jev-paint-achimala.

ui-generator-instinct-jev: composition from a fixed catalog

ui-generator-instinct-jev (7 stars as of September 2026) turns a UI description into a selection of existing shadcn/ui components, fields and styles. It breaks a request down into a series of choice-and-score questions and maps the answers onto a bounded component catalog, rather than generating new markup or JSX from scratch.

That boundary is the project's actual claim to make: it demonstrates UI composition through decisions, and its own listing is explicit that "Jev itself does not generate page code or copy." Practically, that means the output is only ever as varied as the shadcn/ui catalog you give it — new component types, unusual layouts, or copy generation are out of scope by construction, not by current limitation. At 7 stars this is closer to a proof of concept than a tool ready to replace a design-to-code workflow. Aiskill.market entry: /skills/ui-generator-instinct-jev-joevidev.

What the pattern reveals

Across all three, the constraint is the same: Jev returns typed scores and choices, so any "creative" output has to be reframed as picking from a bounded set — sentences scored against a rubric, pixels drawn from a probability distribution, or components chosen from a catalog. None of these projects asks Jev to write a caption, design a novel layout, or compose an image from nothing, because that's not what a typed-decision model does. If you came to this series expecting Jev to be a small, cheap image or text generator, these three projects are the clearest evidence that it isn't one — see what is Jev / TypeSafe System One and typed decisions vs free-text agents for the underlying distinction.

What to do next

1. Don't treat jevmeter's scores as ground truth. Use them as one rubric-based read of a transcript, and check them against a human pass before publishing anything scored automatically.
2. Run jev-paint's local relay before assuming it's broken — the CORS restriction means the frontend genuinely cannot call Jev directly without it.
3. Keep expectations bounded for ui-generator-instinct-jev. It's a component selector, not a layout generator; feed it a catalog that already covers what you want to build.
4. Read the "About Jev" framing on each project page before assuming any Jev-branded tool generates content — most don't, by design.
5. Compare against a conventional tool in each category (a real video analytics tool, a diffusion image generator, a design-to-code assistant) before adopting one of these for anything beyond exploration. The Awesome Jev radar is the best place to watch for more mature entries in this creative cluster.