web4kit-starter
web4kit starter site whose home page is planned per visitor by a System One model (Jev or local LAYA), with calibrated fallback to rules.
Overview
web4kit-starter is a generated starter for web4kit, where pages are not laid out by hand. For each visitor, a System One decision model decides what to show, with which component and where, based only on the visitor's situation. Preview personas such as first-visit, from-instagram and night-owl show how the page changes.
Three engines are available: hand-written rules (offline, free and the fallback whenever a model is down or uncalibrated), LAYA running in-process (free and offline, about 1.7 GB of weights and about 10 s per new situation on a laptop CPU, then cached), and hosted Jev (about 300 ms and $0.0002 per new situation, per the README). In production, an engine's answers are only trusted where its calibration profile says so; elsewhere rules decide. Development mode shows ungated answers and an X-ray overlay of each decision, and a check command fails CI when calibration is stale.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | What content should be shown to this visitor? | — |
| Choice | Which component should render it? | — |
| Choice | Where on the page should it go? | — |
How it works
- A visitor arrives; their situation is described.
- The engine (Jev, LAYA or rules) decides what, how and where.
- Answers are used only where calibration allows; otherwise rules decide.
- Results are cached per situation.
Strengths
- Calibration-gated model use with rules fallback.
- Free local development with LAYA.
- X-ray view explains each decision.
Caveats
- Starter template; regenerated from web4kit releases.
- LAYA is slow on CPU for new situations.
Getting started
pnpm install && pnpm devAt a glance
- Type
- Project
- Status
- beta
- Language
- TypeScript
- Runtime
- web app (Node)
- Has code
- yes
Use cases and integrations
How curated lists describe it
Standalone web4kit starter: pages planned per visitor by System One models.
wh000wh000/awesome-jev-live
Frequently asked questions
What decides the page layout?
A System One model chooses content, component and placement per visitor situation.
What if the model is uncalibrated?
In production, hand-written rules decide instead.
What does Jev cost here?
About $0.0002 per new situation, per the README.
Related projects
React components that use Jev judgments to choose which component renders, how lists are ordered and whether affordances show.
ProjectbetaTypeScriptPrompt-first music studio where Jev makes every compositional choice, from plan to each bar pattern, rendered as editable MIDI.
ProjectbetaTypeScriptVercel Labs' generative UI framework with an experimental compose path where Jev picks which catalog components and actions to render.
ServicebetaTypeScriptCommunity shadcn-style React components and interactive lab for building TypeSafe AI interfaces.
ProjectbetaTypeScriptEmoji autocomplete that asks Jev eight parallel questions per typing pause: emoji choice, emotion, mood, urgency, sarcasm and more.
ServiceprototypeTypeScriptDrop a file and get a viewer: Jev composes UIs for known formats and routes unknown ones, with Haiku inventing specs only when needed.
ProjectprototypeTypeScript