Jev for files and documents — a daily feed

https://t.co/78HxuqZh57 自作のPDF分析ツールにJevを組み込んでみたときの話

3/4 A partir del texto jev resuelve varios if de preguntas dicotómicas (Noul) y de choices . Cuando se activan esos if, busca en mis datos de natación, agrupa, se cambia al panel de comparación y lo muestra https://t.co/HrDoYaggII

האפליקציה החדשה שלנו, Replaceme, נמצאת בפרודקט האנט... סוכן חינמי מבוסס Jev שיושב על המק שלכם ושולח סיכום של הנוטיפיקציות החשובות באיימסג׳. וזו רק ההתחלה כמובן, יעשה עוד הרבה בהמשך. נסו! https://t.co/oQywtjgvuW

Jev + Claude is the reason I can process 40 page technical papers without spending an hour fact checking the output Most people paste a massive document, PDF or article into a prompt and ask to "summarize it". The model gets overwhelmed by context, skips nuance and spits out robotic bullet points I fixed this by putting Jev in front of the reading loop: Raw document → Jev extracts atomic facts → C

一期さんの想起システムも内側では色んな判定式を通して、「今本丸の事を気にかけるシーンか」まで判定して記憶を渡してます〜 男士たちの情報も全部渡せないので、思い出すときに必要な子のだけ、とか。 カスタムjevで思い出してから発言するまで大体7秒くらいかなー結構速いです🤭 https://t.co/rHJ4L4k39Z

Jev and Opus 5.5 were pointed at a folder of signed contracts, and the system flagged what none of them contain. Reading a document for what it says is the easy half. This one builds the list of clauses a contract of that type should have, then checks which are missing. > BUILD - Jev assembles the expected clause set from thousands of agreements of the same class > MATCH - each signed contract map

Stop collecting agent examples by blog post. Start with the decision. awesome-jev is a curated index of public Jev examples for AI engineers studying decision-model patterns. It helps you find implementation paths faster by organizing examples around the decision being made—not the blog or project that mentioned them. Key features: • Decision-first index – browse patterns including tool selection,

I gave my second brain a brain scan. jev-the-janitor 0.5.8 is open source, for Obsidian and markdown vaults, powered by @typesafeai Jev. Each note is redacted locally, Jev answers typed questions about it, and code decides what happens. https://t.co/RpZuaq7d65

Built a personal AI agent called Cuckoo. It can browse the internet, code, use connectors, and I’m currently using it for job applications. The interesting part: Cuckoo automatically organizes your chats and projects using a jev classificatoni model https://t.co/LOU86QgRJC

I've been having super fun with @typesafeai Jev, burning through 105m tokens ($4.23, amazingly) so far. I figured it could be pretty great plugged into Excel, so I created an add-in (Open Source). Check it out! https://t.co/ED62wH6Z8Z https://t.co/BcLOhNx74l

A 2017 ICO LANDING PAGE WITH A FAKE COUNTDOWN BECAME A 2026 PRODUCT SITE IN 58 SECONDS. JEV DECIDED WHAT GOT DELETED. the clip is 20 seconds. Opus 5.5 rewriting a page on the right, a terminal full of Jev verdicts on the left, and a before and after at the end. 1 prompt → 6 sections → 41 Jev decisions → +1,642 lines → 58 seconds "The 2017 token sale closed long ago. What's left is the network." -

@hwchase17 jev handling all the small decisions inside a langgraph graph is a combo we keep seeing. we catalog what people build with jev at https://t.co/XaUVtU9OPx - would love a submission

I built https://t.co/uebcvfseUA, a directory of projects built with Jev. I don't pick what goes on it. Every 6 hours a cron pulls candidates from GitHub and HN, and Jev judges each one. 1,151 judged so far. 500 made it in. https://t.co/doMENRHHa5

What the Jev #02: @jackcheng points at a shape, says "put that there," and it moves. No vision model. Jev reads the canvas as text and answers 8 multiple-choice questions. https://t.co/mSCirvd2C2

Combined design system atoms, Jev and Opus to compose custom screens based on user input with dynamic renderer shaping own experience https://t.co/s0PfDGDoCy

We turned TypeSafe Jev into a real-time feed scanner for X. It scores every post on your timeline in under 120ms and marks likely AI slop before you read it. Over 80% confidence gets a red boundary box and a loud badge. Under 40% gets clean green. Nothing is filtered, hidden, or collapsed. Full extension repo and demo in the first comment.

Built a random demo for the event at @betaworks to explore of @typesafeai new model Jev. Sorted a large an amount of files into a file system using serial classification questions based on file name and type and if unsure progressive disclosure of contents. https://t.co/NHMkrOQGct

After 341 calls to Jev, a judgment-only model, inside my knowledge vault, the question that matters isn't "how accurate is it?" It's "where exactly does it go?" In knowledge management I see two immediate places: routing notes into categories, and cutting tokens before expensive models. Same principle for both: make the judgment once, cheaply, and put execution and expensive generation behind it.

I ran @typesafeai on a dataset with 68,912 rows here's what we found! Only 28% of finance-alpaca is actually about finance. 68,912 rows. 19,047 are finance. The rest is general Alpaca. I cleaned it for ~$1–2 with a cheap classifier (TypeSafe's Jev), then fine-tuned a small model on it. It didn't get measurably better. What did change:

I've gathered 1000+ Typesafe/Jev replications, models, interpretations and papers in one dataroom, everything since the 9/15 release. It's wild how hyped people are about general-purpose discriminative models & classification in late 2026. You can browse the dataroom or download it as JSONL/CSV, throw it to your autoresearch and start hill-climbing from there.

競馬の論文 100 本を Jev で仕分けて、LLM と速度とコストを比べた https://t.co/VRf5Qlif9Q

everyone building a "curated directory of Jev projects" and nobody building the "curated directory of curated directories of Jev projects". 🤪 anyway I fixed it → https://t.co/Ab7s1El9Nt https://t.co/Hz26wsjgAt

Yes, this is one of those "let's connect" posts. Hi, I'm Shrithan 👋 AI engineer from Bengaluru. Until recently I built production AI agents for a US startup. Now I have a lot of free time and absolutely zero chill. So in the last 10 days I: → had JEV read all 464,720 AI papers on arXiv → made every YC company "interview" my resume → graded 25,784 CLAUDE.md files → checked if Anthropic really fixed

I finally found something useful to build with Jev. Meet Clearbook. A financial statement analyzer powered by Jev. https://t.co/zQw8A3dIUL https://t.co/ReyESwf2OB https://t.co/zKVTwwNNPP

Stop paying a chatbot to think out loud while you still do the job. A reply is not a result. A result is the brief, the comparison, the recommendation you can send the same day. Jev + Nova is the desk that does that work. You drop one prompt. Nova opens 5 sources, reads the official docs, and writes one verified brief. Jev watches the confidence. Above 0.80 it ships. Below 0.80 it asks you before

Cheap, fast, and efficient OCR + RAG pipeline with @vlmrun Gateway and @typesafeai Jev. https://t.co/jvyLsAaJHr

hoje em dia tem muita coisa saindo e acompanhar tudo tava ficando impossível pra mim montei um digest diário de engenharia e tech. rss o dia inteiro → o jev classifica → um score em python escolhe o que é relevante para receber as próximas edições: https://t.co/TXD9EDQw0J

エージェントがファイル編集する際にJevでチェックを掛けるPi Agent用プラグインできた。 - 対象ファイル毎に任意のチェック項目を追加可能 - 項目ごとに閾値を設定可能 当分これで満足できそう。 https://t.co/rJBQPqf5bo

Jev、これ普通にヤバい。 「AIは税務書類の分類が苦手」 と言われていたのに、 税務書類を100%分類。 しかもLLMより34倍安く、6倍速い。 Jevを使って「税務書類の自動分類システム」を構築。 これまでは、昨年作ったLLMパイプラインを使って、数千件の税務書類を処理していた。 今年4月頃まで「AIは税務書類の分類に失敗する」という記事もいくつか出ていたけど、今回Jevで改めて検証。 結果、 ・税務書類のコーパスを100%分類 ・コストは1ページわずか$0.001 ・従来のLLM構成より34倍安い ・さらに6倍高速 しかも、この税務書類分類システムはオープンソース化。 大量の書類を「LLMに全部読ませる」のではなく、 Jevで高速・低コストに処理する。 こういう事例を見ると、Jevの使い道はまだまだ広がりそう。

#rebuildfm RAW版（編集前）の文字起こしから自動でチャプターつくるのも jev でまずセグメントを分割（ここでトピックが変わったか？をタイムスタンプごとに判定）してそのあと Claude Code のサブエージェントに投げてチャプタータイトルつくる、でできた。 jev の部分は$0.01, 0.5s https://t.co/k7lobZhc4f

I like this approach to RAG with Jev. For each passage, Jev returns 4 probabilities like: P(relevant) < 0.45 → drop P(has answer) > 0.55 → keep It keeps or drops based on probability! https://t.co/Xd5uOKa9Uw I've been trying it out for compaction in copilot sdk... https://t.co/a2gcnzFt4g

Kindleで出しました。 『「Jev」入門 ― 文章を書かないAIが、問い合わせを3つに振り分ける』 ・3つの問いの型と「確信度」の読み方 ・確信度0.9/0.5で「自動・確認・人」に分ける線引き ・計算や日付が苦手、など向かない仕事 フルカラー漫画30ページ収録。700円／Kindle Unlimited対応。 https://t.co/3o0981x1mK

Everyone is using Jev for data but have you tried it for context? Only anecdotal so far but it is a very powerful tool to add to the toolbox. Best approach is LLM + Jev get the agent to loop to improve Jev. From big dawg Opus 5.5 (I actually like this new one): --- The breakthrough, stated precisely: the agent turns its understanding into a plain-English question once. Jev runs that question over

11485条数据，jev 1分钟跑完了，确实震撼🥲 https://t.co/jsd8Wrrizi

Ran Jev by @typesafeai + #Docling + #Gemini 3.8 Flash. I evaluated document processing use-case on these 3 combinations. 👉 Docling → Jev: 2.0s in $0.0021. 👉 Docling → Gemini 3.8 Flash: 10.9s in $0.013. 👉 Gemini (End2End): 11.9s in $0.020. https://t.co/oFJiiZ7Hp9

Jev doesn't chat but we can ask it about events. Fable suggested checking Jev against movies that came out so we've checked it against 880 movies we were able to see a hint of cutoff being in the beginning of 2025 where it was less than 50% sure that the movie came out which in fact came out