Can the hosted demo judge my own cancellation clause or photo?

No. The browser runs recorded observations for fictional holidays plus real application rules, so changing budgets or photo selections changes the calculation but uploading your own text or photo does nothing.

If I run the training recipe, do I get the model from the video?

No. The recipe trains a separate educational ModernBERT classifier from a small synthetic set, and the guide states it does not reproduce the video's DeBERTa model or its score.

Does the eight-decision baseline and train loop prove useful accuracy?

It only proves the pipeline ran and the expected artifacts exist. Eight decisions are an explicit smoke check, and the bundled train, development, and test scenarios are synthetic examples, not accuracy evidence.

What setup should I expect for the demo versus the training path?

The sample app needs Git, Node 22+, and npm with no model setup, while training needs Python 3.12, pinned dependencies, storage for the base-model download, and enough memory for the model.

How do I run the sample app?

Clone the repo, cd into apps/public-demo, run npm ci then npm run dev, and open the local Vite URL; it needs Git, Node 22+, and npm, with no Python or model setup.

What is the smallest training loop?

From the repo root, uv sync --frozen under apps/agency, then run tools/tutorial.py prepare, download, baseline --limit 8, train --limit 8 --epochs 1, and test, which writes a checkpoint, baseline, freeze, and test artifacts under workshops/.

How do I adapt the pattern to my own work?

Start with one input, one question, and three precise answers (for example a cancellation clause, whether a cash refund is available, and yes/no/can't tell), then check the label rules, measure the original model, freeze a checkpoint on development results, and open the final test.

CommunityJev Demos & Experiments

Away Together Starter

Free starter that teaches turning one document into one small typed decision, built around a fictional travel agency where recorded observations meet inspectable application rules.

See highlights

Open github.com

Stars: 13
Language: HTML
License: MIT
Published by: Community
Author: earlyaidopters
Last verified: Oct 2026

earlyaidopters/away-together-starter

Organization repository on GitHub

View on GitHub

Free Jev-inspired AI starter: visual guide, recorded travel demo, reusable prompt and small training recipe.

Forks: 4
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.

About this project

A free, runnable starting point from Mark Kashef and Early AI Adopters. Turn a document into a small, useful decision. See how it works, then train your own. Try the demo · Follow the visual guide · Copy the full prompt. A hotel offers “flexible cancellation.” Does that mean cash back, hotel credit, or not enough information?

Quoted from the project's README. Read the full README.

Tagscommunity jevbest catalog-import html

Works forJev for Classification

Used inJev for Software & DevTools

Highlights

Interactive travel-agency demo where excluding one photo flips a traveller from decline to needs review.
Recorded-observations design that preserves uncertainty instead of turning missing evidence into a yes.
Smallest complete loop from prepare through baseline, train, and test with frozen runs that refuse overwrites.
Prompt and guide for adapting the pattern to one input, one question, and three precise answers.

Watch out

The hosted demo uses recorded observations for fictional holidays and cannot analyze new text or photos; the training path builds a separate educational classifier and does not reproduce the video model or its score. Copy reviewed against the repository README on 2026-10-04, not from running it.

Questions about this entry

Short answers drawn from the linked project or page.

More Jev resources like this

OpenJev

Python decision client exposing Choice, Noul, and Score primitives with isolated order-invariant scoring, a deliberative flywheel that crystallizes hard cases into fast paths, and local or hybrid deployment.