I made this financial demo to explore how an interface could start matching a request while someone is still typing. Jev can identify a suitable view before submission, and a preview shows how it understood the question.
Try the live prototype or watch the full walkthrough. All financial data is fictional.
From language to a view
Jev returns typed decisions: choices from options the application defines, with probabilities attached. TypeSafe describes the model here. Its output can say that a question asks for a share count, for example, without having to express that choice in a paragraph the app then needs to interpret.
In this prototype, Jev works out the investment, period, and calculation as the request takes shape. Once that interpretation is ready, Enter confirms the match. The app checks the choices, computes the figures from its records, and selects the corresponding view. GPT-5 adds an explanation when one would help. A straightforward lookup can stop at the figures.
That division gives each part a smaller job. Jev handles the wording of the question. Code owns the arithmetic and presentation. The Thinking panel lets someone inspect the interpretation and its confidence.
Walkthrough
The useful part happens while I type. Jev can identify the request before I submit it, leaving the app to apply a known view.
Read the walkthrough or jump to a step
- Current step
Jev starts interpreting the request during typing, so it can match a known view before submission.
The preview shows how the app understands the request, giving someone a chance to spot a mismatch before committing.
Enter reuses the matching interpretation. Work that began during typing carries through, without asking Jev to start over.
The app fills a designed component with computed figures. The same validated choices and records produce the same structure and numbers.
The log makes the interpretation inspectable. A wrong answer can be traced back to the choices that shaped it.
How is my money spread across accounts?
Send $200 to my cash reserve
What stays predictable
Given the same validated choices and the same records, the app produces the same answer structure and calculated figures. That is where determinism enters this version of generative UI. Jev's interpretation remains a model decision, which can vary or be wrong. A confidence score does not remove that uncertainty, and an optional written explanation can vary too.
I narrowed an earlier approach that gave models more freedom to arrange the page. Defining the views myself made their behavior easier to inspect, but it also means each new kind of answer needs design and implementation work.
I could see this approach extending to a larger set of tested components. Different ways of asking a financial comparison might lead to the same layout, with matching units and a shared scale. Interpreting the request during typing could also reduce the pause someone feels after submitting it. The language could stay flexible while the resulting interface follows known rules.
That would give designers a way to work on generative UI through the options it can choose and the conditions for choosing them. This demo only covers a limited set of questions and fictional records. Whether that approach feels useful to customers is still something to test.