Jev can generate interfaces in under 2 seconds. Almost zero cost. I took the shadcn component library... → hooked it to a custom style md file → made some basic UI rules of popular apps using the Mobbin MCP → Jev instantly decides based on the predefined library and... gives you a fully responsive composition in 1-2 seconds. Adding an LLM layer at the end improves output efficiency much better but takes aorund 8-12 seconds to generate (Qwen, Haiku)show more

Cool use case — Jev picking shadcn pieces is sharp. Grabbed it for

🙌

is it technically "generation" tho? more like pick the right UI? 🤔

imagine a design system and a set of rules, it mixes both of them and picks the right thing and can fill in with some source/demo data. calling it jeverative (haha). check the architecture diagram in thread

Tighter and defined rules lead to more refined outputs.

I got it down to 400ms

For small components, without much rules/styles to choose from- it is really fast

Was discussing with @karrrtiiikkk how generative ui flows would be a possibility because of Jev.

Using Mobbin MCP to establish the UI rules is super clever, how granular did you have to make that style file?

for now it just took some top fintech app references, for some complex queries Jev-only mode will throw an error. Jev+LLM might still be able to give something. Also, depending on your style filesize – Qwen has larger error rates but costs very less. Haiku does better outputs but costs more.

@rohanchkrabrty Nice! Was thinking about this an hour back. I felt this would reduce hallucinations too and choose elements from the prescribed design system only.

@rohanchkrabrty yes. more rules and guidelines lead to better output. adding an llm to it is the most idea solution.

@rohanchkrabrty So do you send the rules to jev in every request ?

@rohanchkrabrty not the entire rule set. based on prompt, a local code (or llm, in hybrid mode) does some filtering before passing on to Jev.

@rohanchkrabrty This makes sense. In my today’s exploration, I choose to filter top 60 but it would make sense to add some logic filters prior to sending to Jev. But its definitely not as fast on my local app.

@rohanchkrabrty what kind of logic filters are u passing? written md/schema

@rohanchkrabrty I did like 5min work, not great. Just to explore how reliable Jev’s output is. Right now was just using top 60 tags from my bookmarking tag list. Ideally should send most relevant top tags for evaluation.

Same shape here, filling widgets instead of components. Two things bit me: options dilute (leader 0.99 -> 0.81 as I added them), and one extra sentence in a question dropped accuracy from 92% to 68%. Shortlist before Jev sees the library.

Really neat demo!

🙌

boy is fast

The style system is the real leverage

yes! Tighter rules lead to consistent outputs

Exactly, less variance is the point

The speed is impressive, but the style rules are doing more work than the model call. The useful next test is whether the generated interface stays coherent after several edits, not just whether the first screen looks right.

It can create some variations and live targetted edits. But prolly needs an LLM to do a lot of handholding

You Sir, definitely need to do a closed door masterclass.

👀👀

jeverative 🤣

This is super cool! Had tried something similar with the components and slots for pages defined upfront, got really promising results, check it out :D

Reminds me of this. This one uses pure json schema shuffling (sensless shuffles)

Jev with component library is a good direction.

reallly ? it's awesome

It chooses from pre made components... its not really building something from scratch

that is the whole point of a model like Jev. it chooses fast, coz it doesnt have to build from scratch. And isnt it why libraries and DS exists in the first place?

did you put this on github ?

yes its private repo as of now.

how does Jev handle edge cases or unique design requirements?

It simply cannot until there is a rule. Hence added a Jev+LLM composition which takes care of these cases

@grok what is jev

Can you share the code?

jv

Generative ui is two tasks - choosing components + hierarchy and populating props - how are you able to do 2nd with JEV? first one is straightforward but getting JEV to fill-in values of props - is it JEV-only or llm layer does the population?

it populates from a mock data source in this demo. It needs a local code or llm to help do its job effectively

Generating code at sub-second speed is trivial. The real cost is the inevitable accumulation of 'generated technical debt' that human maintainers have to untangle six months later.

UPI पर MDR चार्ज का असर व्यापारियों और ग्राहकों पर क्या हो सकता है? इस विषय एक छोटा सा इंटरव्यू जरूर देखिए और अपनी राय साझा कीजिए। 🙏 👇 पूरी बात यहाँ 🙏👆