All builds

Open-model Jev-style server that reads next-token logits to return typed choices, rubric scores, and truth judgments.

Pi extension with a batched Jev evaluation tool, terminal playground, and typed API for other extensions.

Text linter that uses Jev Score judgments to evaluate strings embedded in source code against configurable writing rules.

CLI that uses Jev task classification to choose and launch Cursor, Claude Code, Codex, or OpenCode with an appropriate model and effort level.

Claude Code and Codex hooks that ask Jev which parts of long shell output remain relevant before sending them into model context.

ACP and MCP adapter that exposes Jev typed decisions and computer-use actions to Codex, Claude, OpenCode, and other agents.

Jev-compatible API server built on open models and SGLang, with prefill-only inference and typed decision endpoints.

Trainable family of small Qwen-based Jev-like decision models with typed primitives, datasets, evaluation tools, and local inference.

Fast natural-language search for Azerbaijani names, built with Jev. 💬 "a Turkish girl's name, flower-related, not too popular" → ranked results in half a second 11,000 names, 93 semantic traits. Unlike an LLM, Jev doesn't write answers - it makes decisions. ~0.5s. $0.00025 a query. #jev #typesafe https://t.co/DaHVDrdeQ1

Meet JevvieScout, built with Jev😺 It watches posts and tells you who’s worth DMing based on your goal https://t.co/CE3ItIAfPv

I built a thing using Jev. Have you ever been annoyed by not finding an old Claude Code or Codex chat? => Now you can just describe what it was about and Jev finds it in milliseconds. Open Source: https://t.co/RbAgyiZWeO https://t.co/ltuW7VNKEk

I gave 12 residents the ability to gossip. Whisper “Free Coffee in cafe for next hour” to one person. Watch the 12 resident rush to the cafe shop. Built with Jev. Try your own rumor: https://t.co/KsNhS9DG5U https://t.co/Io8KfbNKKa

i genuinely don't understand why anyone is using combo of "Jev + Polymarket" as trading router. i just built jev layer with agenkit in my trading system that gave me edge to create alpha. i am openly leaking the cheatsheet. Bookmark before it's too late and start using Jev in your trading system.

Been experimenting with using jev as an attention layer over code retrieval. On 356 held-out SWE-Explore tasks, adding Jev reranking to the same 100 zvec candidates improved core recall@1k lines from 11.96% → 22.95%. More here: https://t.co/ZNp3AnD8Bn #jev #TypeScript

Jev the Band: I made a jam band using jev. (🎧on) I taught jev how to read and write music Guitar, Bass, Drums, and Keys controlled by jev lights controlled by Jev soundboard, yup, it's Jev I've been obsessed with recording Jev's jams this weekend and now I'm sharing them with you

A fully autonomous real-time trading bot built with Jev in one evening and morning, ingesting onchain and offchain data for rapid decisions, has lost $31,680 so far. https://t.co/lYAShIq9cg

A tiny open source browser agent using Jev instead of an LLM for every click. Found a flight search in 7 SECONDS. Total cost: $0.0039. Here's why that's not a typo. A normal browser agent asks a chat model "what should I click" on every single step. That's a full generation call, just to pick a button. Mine doesn't. The DOM state at each step becomes the input. Jev gets the available actions as a typed choice question. It picks the action, not by generating text, by classifying against what's actually on the page. The only place a language model still runs is typing free text into a fie

Semantic code search CLI and MCP server: Jev scores authorized repository excerpts and returns exact file paths and line numbers to coding agents.

Decision plugin for DeepSeek Harness and iPolloWork that lets OpenCode, Codex, and other agents use Jev to select tools, skills, and task owners and evaluate outputs.

Rails-native Jev wrapper that turns typed, calibrated decisions into application control flow.

Open local System One implementation for NVIDIA DGX Spark, with a Qwen3-based model, Jev-compatible API, architecture notes, and evaluation tooling.

Community Go client for the TypeSafe System One API with typed question builders, retries, examples, and jev-latest as the default model.

Anyone else playing with JEV? Feels like that’s all I’m seeing today 😅 I built a small PoC using JEV as a pre-screening step for reverse engineering, before sending the APKs to an LLM for deeper analysis. The flow is pretty simple: `APK → static analysis + Quark → JEV → score → reverse or skip` The goal is to avoid burning LLM tokens on samples that don’t really warrant deeper reversing. It’s still early, but the token savings are already pretty interesting when you’re triaging a bunch of samples. #Reversing #LLM #JEV #Malware

Okay so we also built something else. So we are using @Muse & @typesafeai Jev together now. I’m using Jev to grade my AI assistant Sarah on her own work before I ever see it. So (Jev by Typesafe) will act as the judge. Every caption draft gets scored against my actual writing voice: my hooks, my outro style, my banned words, all of it. And It's bigger than captions. Before she states a fact, the judge checks it against her own records, it caught a wrong number tonight before I saw it. Before she acts on a vague instruction, the judge decides: act, or ask me first. A checklist that never

using Jev to score 2,952 museum objects on "is this architecture?" https://t.co/GaJgJtMvdb

I found the "Internet" moment in using Jev Jev analyzed every LLM and chose the BEST-VALUE one for each prompt It cut my costs and time by ~80% - and it’s probably the best way to use Jev... here's how to setup it in 10 min: step 1 → open @typesafeai , create API key (keep it off chat paste) step 2 → store TYPESAFE_API_KEY in a local .env step 3 → clone the Github below, then npm install && npm test && npm run demo step 4 → test Jev before connecting an agent: npx tsx sdk-runner/cli.ts --route-only "Rename a button label" step 5 → dry-run first: npx tsx sdk-runner/cli.ts --dry-run "you

i've built an extension for twitter that ranks every tweet by relevance based on your interests using @typesafeai 's Jev, highlights very recent tweets to maximize engagement and assesses your tweets based on your goal (also using Jev) should I open source it? https://t.co/GgQjTT2NWy

I wanted to post a preview of our Jev benchmark dropping tomorrow. We are using Jev to generate symbolic programs (state machines) through an iterative process. The results so far have been amazing! The old process took ~2–10 seconds. With Jev, it’s ~0.3–2 seconds. The costs have also dropped by an order of magnitude. Jev is also used in control flow, removing the need for brittle heuristics.

I built a token compression engine using Jev. It evaluates document (huge text) chunks in parallel, in sub-100ms passes, stripping 85% of boilerplate so we only pay Claude or GPT-4 for high-signal answers. Here is a demo https://t.co/3nVjtG8zZY

noticed below cumulative impact of using Jev + Jevcache + Fable: - Per-decision: ~220–780ms Jev vs 20–40s claude before - Loop's decisions are ~50× cheaper; Fable was consulted only on the 3 low-confidence steps - jevcache: earlier identical-phrasing re-run showed cached (0ms replay) live.. repeat run was 26% faster (7.2→5.3 min)

Jev can outperform the rules many companies use to prevent fraud. We found that without any pre-training, it accurately detected 93% of a fraud ring. In contrast, an LLM in a similar set up only achieved 62%. https://t.co/ronCD7hLvb

3 days. First AI product I've ever built, with Jev doing some of the heavy lifting inside it. Here's the landing page. No Figma at all, just Cursor ai and the design skill I built. Launching soon https://t.co/VBhWBCVEMf

I just classified a month of my Gmail for eight hundredths of a cent. 150 emails → 9 folders. Read-only, nothing written back to my inbox. Built with Jev. Repo below 👇 https://t.co/PaUcxpcDE1

I builded Jev Reflex which can control your mac with hand gestures and voice. It is build with Jev 1.13 from @typesafeai . I builded it with @OpenAI Codex using Astra and Sol. It is using Jev credentials from @OpenRouter and for all this just used $0.01. As you can see in the video, it recognizes hand gestures and voice and can do various task on mack, like mazimize or minimize anything you point you index finger. And then pinch gesture to complete it. It took me 2 hours to build it on a monday morning, which included code, enabling controls on mac and calibrating it first. Project availbl

Using JEV by @typesafeai to classify the 100s of inbound deals we get, side-by-side with a small open model we’ve been using for our in-house AI rating module that self-trains. impressive how close JEV gets while being significantly cheaper and faster to run for classification :) https://t.co/6c6I9SHQuY

There's a fun mix of excitement and cope on my timeline re-Jev: - OMG I made Jev a year ago this is dumb - You can fine tune this encoder into a better, faster, local classifier, etc Re the "just train your own small classifier" crew - Voyage trains great, SOTA pointwise rerankers that are obviously specialized to the reranking task and have been at it for years, likely also using a strong synthetic data recipe... Exact setup in the response tweet. Do note that using Jev is definitely more expensive than using Voyage rerank-3 bc of the tokens in the lengthy rubric included in every request (t

Use @typesafeai's Jev model with Pinecone to rerank results with natural language criteria! Usually with rerankers, it's hard to cleanly specify what should and shouldn't be returned in results. Jev resolves this by refactoring the problem into evaluating against distinct binary criteria, which pairs great with Pinecone retrieval! In this demo, we compare using Jev and Claude to rerank 200 returned candidates from Pinecone. Jev returns a reranked list in about a second — 830 to 1,300 ms across eight test queries. Claude Opus 5, doing the same job in one long-context call, takes 4.2 to 6.8

First real use case with Jev got up and running last night. Used it with GLiNER 2.5 base to help dramatically speed up entity / relationship extraction. GLiNER by itself for Graph RAG over a huge document corpus is fairly unreliable. Using Jev to help determine what it got right and then send everything else to my local 35B model. ~25x performance boost.

simple real-time bug checker made with jev in case you code by hand sometimes also works with english (underlines inaccurate sentences) https://t.co/hJt7xzRLpB https://t.co/XsitOtjnnh

We just cracked how to make a great Meta ad using Jev ( a classifier model from TypeSafe AI. ) We pulled around 2,000 real ads from 8 beauty brands off Meta’s Ad Library and graded 302 of them with Jev across 9 dimensions: hook, format, funnel stage, CTA, claim risk and boldness. Jev is an AI evaluation tool from TypeSafe AI that lets you run the same structured review across hundreds of inputs. The best ad in the entire set was a drag performer climbing 300 stairs to see if her makeup survives, for e.l.f. Cosmetics. She puts the makeup to the test by climbing the stairs, and that’s the ad.