0 deps no server offline private

Run typed-decision models in the browser. One state text plus any number of typed questions (choice / score / noul) go in; a calibrated probability distribution per question comes back from a single forward pass.

▶ Try it live below npm GitHub 
import { createDecider } from "jev-web";
const decider = await createDecider();
const { answers } = await decider.decide("I was charged twice for the same order.",
  [{ type: "noul", instructions: "The customer is asking for a refund." }]);
console.log(answers[0].noul, answers[0].probabilities); // 0.92 { no: 0.08, yes: 0.92 }
—
latency, warm
—
forward passes
—
questions answered
$0
cost per call
—
network calls when warm
last 40 runs

Input

model not loaded
Questions — all answered in one pass

Typed decisions jev-web · one forward pass

—

Load the model, then press Decide. Editing the state re-runs automatically.

Raw text parsing regex / keyword rules

—

The other way to do this: a hand-written rule per question.

The baseline is a fair keyword matcher — it wins on easy text and collapses on negation, contrast and mixed-topic states. It is also N separate rule sets to maintain, and it can only ever return a hard label: no distribution, no confidence, no way to threshold it.

Network panel proof, not a promise

online
Requests to load the weights
—
one-time, cached by the browser afterwards
Requests while answering (warm)
—
press Decide to measure
Content leaving this tab
nothing
the state text is never transmitted
outbound hosts observed by the Performance API
none yet

jev-web vs. calling a hosted LLM API with fetch()

jev-web (this demo)hosted LLM API via fetch()
Cost per decision $0.00 — weights are static files per-token, forever; a busy form multiplies it by every visitor
Latency measured above — all questions in one pass typically 0.4–3 s, plus a network round trip
Output calibrated distribution per question, every question in one pass prose you then have to regex back into a distribution
Determinism same input → same probabilities, always temperature 0 + structured outputs helps, still not guaranteed
Offline works in airplane mode once cached dead
Privacy state text never leaves the device your users' text is sent to a third party on every call
Ops burden static files on a CDN, no API key, no rate limit key management, rate limits, quotas, uptime, spend alerts
Ceiling only picks among options you supply — never writes, never explains general purpose; will happily hallucinate an option you never offered

The honest trade: jev-web is a classifier, not a chatbot. If you need free-form answers, you need a generator. If you need a routing decision, a score, or a yes/no with a confidence you can threshold on — this is the cheaper, faster, private one.

Install

npm i jev-web @huggingface/transformers

@huggingface/transformers is an optional peer dependency (imported lazily); hosts that already ship their own copy can inject it via createDecider({ transformers }). This page runs on GitHub Pages, which cannot set COOP/COEP, so the ONNX runtime stays single-threaded; run npm run dev locally for the threaded WASM build.