Are the Jev-versus-Gemini speed gaps real?

Only for that run. The repo says results depend on momentary load, model, and prompts, with Gemini varying from seconds to over 30s, so the ratio also varies.

Can I deploy this for real fraud blocking?

No. It is a simulation for local or test use, not for production or real fraud decisions, and each evaluation burns tokens in both models.

What happens when Gemini hits quota?

A breaker switches to contingency labeled MODO SIMULADO, which is excluded from comparisons; Jev still returns typed decision and fraud probability.

How do I run it without leaking keys?

Clone, npm install, copy .env.example to .env with GEMINI_API_KEY, JEV_API_URL, and JEV_API_KEY, keep .env ignored, and use spending limits.

What does Jev return for each transaction?

A typed decision of APROBAR_DIRECTO, SOLICITAR_2FA, or BLOQUEO_PREVENTIVO with its confidence and the probability of each option, plus is_fraud, a noul probability between 0 and 1. No text is generated.

What stack and runtime does the demo use?

React 19, TypeScript, Tailwind CSS 4, Vite, Express, tsx, and @google/genai. It needs Node.js 20 or higher, runs with npm run dev, and serves on http://localhost:3000. The license is MIT.

Where do the API keys live if I skip the .env file?

The app's Settings panel accepts the keys, and in that case they are stored in the browser's localStorage rather than the environment. The .env file is gitignored, and the README warns never to commit keys to the repository.

CommunityJev Benchmarks & Evaluations

System1 Fraud Interceptor Demo

Local demo that runs Jev and Gemini in parallel over synthetic transactions, comparing decision, latency, tokens, and audit record.

See highlights

Open github.com

Stars: 0
Language: TypeScript
License: MIT
Published by: Community
Author: ordepas
Last verified: Oct 2026

ordepas/system1-fraud-interceptor-demo

User repository on GitHub

View on GitHub

Demo de un interceptor de fraude simulado que compara en paralelo un modelo Sistema 1 (Jev, TypeSafe AI) con un LLM (Gemini) sobre transacciones sintéticas: velocidad, costo por llamada y decisiones. Proyecto personal de experimentación, no es un benchmark.

Forks: 0
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.

About this project

Demo de un interceptor de fraude simulado que compara, en paralelo y sobre transacciones sintéticas, dos formas de decidir: Sistema 1: Jev (TypeSafe AI). Modelo rápido que responde con valores tipados (una opción, una probabilidad), sin generar texto. Sistema 2: Gemini 3.1 Flash Lite. LLM que razona, decide y redacta un informe. La idea viene de Kahneman (Pensar rápido, pensar despacio): un modelo rápido para la decisión inmediata y uno deliberativo que la audita cuando hace falta.

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

Tagscommunity github

Works forJev for Evaluation

Used inJev for Software & DevTools

Languages: TypeScript

Highlights

Jev returns APROBAR_DIRECTO, SOLICITAR_2FA, or BLOQUEO_PREVENTIVO with confidence plus noul fraud probability.
Gemini returns decision, natural-language report, and token use; 429 quota trips simulated mode excluded from compares.
All names, shops, and amounts are synthetic; speed and cost figures for Jev are provider own tests.
Runs with Node 20-plus via npm run dev on port 3000; keys via .env or Config panel stored in browser localStorage.

Watch out

Personal experiment, not a benchmark; backend has no rate limit so never expose with server keys. 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

jev-evals

Rubric-based eval harness that judges LLM and agent outputs with Jev through the Vercel AI Gateway, answering all rubrics for a case in one round trip.

Jev Benchmarks & Evaluations#community#github

Community