JEV Case File — JEV cases

语音日记 + 情绪趋势分析

每天录 1 分钟语音，转文字后做情绪打分，月底生成心情曲线和关键词云。

行程规划：不只是列清单

给定预算和 5 天时间，输出含交通衔接、餐厅预订链接、天气预案的完整行程，可一键导出 PDF。

自托管代码审查机器人

接入 GitHub 后每个 PR 自动 review：指出潜在 bug、建议重构，评论风格可以自定义成「毒舌模式」。

给孩子的个性化睡前故事机

输入孩子名字和白天发生的事，生成带插画的定制故事，语气温柔，每晚不重样。

自然语言选股器

「找出过去 5 年 FCF 增速 > 15% 且负债率 < 40% 的消费股」，一句话生成筛选器并回测。

合同审查：风险点自动标红

NDA、采购合同、外包协议丢进去，自动标出不利条款并给出修改建议，法务同事从 3 小时缩到 10 分钟。

批量总结医学文献摘要

喂了 3000 篇 PubMed 摘要，按「研究设计 / 样本量 / 结论可靠性」三栏结构化输出，准确率让合作医生很惊讶。

给独立游戏装上有记忆的 NPC

用 JEV 给 RPG 的每个 NPC 挂了长期记忆，玩家两周前随口说的一句话，NPC 到现在还记得。

一句话分析任意 Excel 文件

丢给它一个 50MB 的 Excel，问「Q3 哪个渠道 ROI 最高」，它自己写代码、跑代码、出图表，全程 8 秒。

语义关键词抽取 + ID 打标全流程

tried something with jev so i tried what you'd expect from me: graph extraction - score every word with semantic significance (1-5) - also tag relevant words with ID

JEV 原生多语言混合对话实测

でん。中央リサーチや種村さんの方式、ファイル名、覚えてる？つかんだボタンを押すと中身を読んでくれる体験……

JEV 已上线 OpenRouter

Jev is already on OpenRouter!! It's only been out a day and TypeSafe still has a waitlist……

把大问题拆成一万个小决策

我们花了 3 年学习如何把提示词写进一个大模型里，而 Typestate 的思路恰恰相反：构建围绕一万个小决策的系统。

JEV：比 Claude 便宜 400 倍

Jev es una nueva IA creada por Diogo Almeida, uno de los creadores de ChatGPT, que promete ser hasta 200 veces más rápida y 400 veces más barata.

不到一小时用 JEV 复刻 Tesla FSD 界面

I rebuilt Tesla Full Self Driving with Jev in less than an hour. This model is a total unlock.

税务文件自动分类器

用 JEV 构建了一个税务文件分类器。我们导入成千上万份税务文件，使用我上一季度搭建的 LLM pipeline……

12 秒给 100 个视频打上销售意向分

用 JEV 对某个 YouTuber 最近 100 个视频做分类：判断每个视频「有多想卖我东西」，并给出销售意向分（1-5）……

Gens, Instinct 与 Meta 的 Crunch 能力实测

用 JEV 跑通了带 Instinct、Ways、Town、Office 的 Crunch 流程……Inworld 的 popularity 场景下需要管理超过 500 人的 NPC 对话。

50,000 Airtable validations for about $30

Validation rules written in language instead of formulas, applied to Airtable: roughly 50,000 checks for about $30, surviving tables being added and removed.

Removing information until the decisions break

A deliberately rigged hiking game: strip away the information JEV gets, step by step, and watch how long its decisions hold up. The author calls it a probe, not a benchmark.

Signal Lab: crypto cohorts scored by JEV

A public page where JEV analysed two crypto cohorts — coins nearing a Top 10 flip, and names reclaiming their SMA50/200 — with setup, ATH potential and narrative per name.

Reranking Pinecone results with plain-language criteria

Pinecone's own walkthrough: use JEV to rerank search results against criteria written in natural language instead of training a reranker.

Support chats: JEV scores, humans label, code decides

A test harness where JEV turns a support conversation into sentiment probabilities, compares them with human labels, and picks where the human-review threshold should sit.

A skill router tested on production chatbot traffic

JEV picks which installed skill answers an incoming request, so the whole catalog stays out of the context window. The author tested it against the routing logic their product already runs.

Talking your day into coloured decisions

A small app transcribes a spoken diary, then JEV reads cues across twelve emotion-linked colours and turns the entry into something structured.

Hooking a deterministic classifier into V8's JIT

An experiment that put a JEV-style classifier inside V8 to help the JIT decide when to promote code. The author's verdict: as dumb as it sounds.

JEVBOT rebuilt its trade filters after a 33% win rate

A live trading bot published its first 45 closed trades, found a 33% win rate, opened the book, audited the failures and rebuilt its entry filters around JEV.

Replacing the LLM guardrail with a probabilistic judge

Running Vercel's safety classifier through JEV instead of a frontier LLM came out 5–18× faster, and a probability cannot hallucinate the way prose can.

Finding old coding-agent chats by description

An open-source VS Code tool: describe what an old Claude Code or Codex session was about and it finds the session in milliseconds.

jevselector: dropping unused skills from the prompt

A CLI tool that scores an agent's installed skills and drops the ones a project never uses — cutting the system prompt by roughly half, while respecting per-agent and pinned skills.

JEV as a Macaly skill for content checks

JEV is now a Macaly skill: ask that app builder for AI classification and it wires JEV in. The demo scores each paragraph for AI slop and flags the writing issues.

Smart mail search and one-click unsubscribe

magic-mail adds natural-language search and automatic classification to the Omamail client, plus one-click unsubscribe.

A Polymarket trading agent built on JEV

An open-source agent that places predictions on Polymarket's five-minute BTC up/down market, with JEV making the call.

Outfit picking as the last JEV check

In the aesty closet app JEV is the final layer before an outfit ships: it caught toe shoes planted in a party look and rebuilt the outfit from clothes the user already owns — 3 s, $0.0015.

Agent step selection in a tenth of a second

Agentcard wired JEV into a shopping agent so it picks each browser step, while a vault fills in the card at checkout — neither the agent nor the browser ever sees it.

A DuckDB extension that classifies rows with JEV

A DuckDB extension that calls JEV directly on any CSV, Parquet file or table — about ten seconds for a thousand rows, no export step and no classifier service to maintain.

Routing multi-agent requests with JEV: 96/97

In a multi-agent workspace something has to decide which teammate takes a request. Fez ran that decision on JEV and scored 96/97 on their own routing battery, 184 ms median, $0.0018 a run.

Pushing a Replit project's 18,360 outcomes down to 57

A Replit project had 18,360 possible branch outcomes and months of manual fixes barely dented the list. JEV was pointed at the project and, per the author, resolved all but 57.