Skip to content

Case wall

Jev AI examples: real builds, updated daily

615 real Jev AI builds and public demos, updated daily. Original text, metrics, and a link back to each source.

Last updated

Public Jev AI examples and demos, not official TypeSafe docs. Clients and SDKs are in the Jev GitHub projects. Paths for jev-latest, Vercel, and OpenRouter are in the September 2026 channel map.

Agent-skill repositories are listed on Jev skills on GitHub.

OpenRouter model ids recorded for Jev AI are on Jev AI on OpenRouter. Rubric and eval posts are on Jev as a judge. How the company and the model fit together is on TypeSafe AI. The pricing calculator uses the published input rate.

Filter

24

quick tip we learned the hard way @ rivet: jev does not reason ...which is by design ...and we should have rtfm first ~ we started by prompt-engineer to ask som…

Guides

Nathan Flurry 🔩 @NathanFlurry

EN

@NathanFlurry shares a tip learned the hard way at Rivet: Jev does not reason, which is by design, and they should have read the manual first. A single prompt that asked for the severity of an issue, with SEV1 through SEV4 as outputs, gave meh results and defeated what makes Jev useful, because an LLM would spend reasoning tokens to reach a conclusion. After reading the manual they split it into small questions: did production go down or become unusable, are they blocked with no workaround, and is something in Rivet broken or failing. Code then returns SEV1 if blocked, SEV2 if down, SEV3 if broken, and otherwise SEV4. He says that works. When a ticket is weighted badly, they give it to an LLM with what they expected, the LLM adds it to a test suite and adjusts the decision logic, and they backtest the new logic on old issues. Within a few tickets the classifier was solid. He calls this a feature of Jev, not a limitation: more control over how each decision is weighed than a generic LLM, and easier to reason about, monitor, and backtest how agents will behave. In their internal support tool the Jev question counts are 13 for support classification and customer status, 6 for severity, 6 for team replies and problem identification, 6 for follow-up timing, 4 for linked dependencies and later conversations, 2 for the addressed teammate, and 2 for conversation grouping. He ends with: Jev is awesome. A screenshot of a real-world classifier is attached.

quick tip we learned the hard way @ rivet: jev does not reason ...which is by design ...and we should have rtfm first ~ we started by prompt-engineer to ask som…

Poster from Treg (@treg_ai)

Apps

Treg @treg_ai

EN

@treg_ai introduces a jobs API powered by Jev from @typesafeai. Try it at https://treg.to/jobs-api. He lists one API to any jobs data, $0.03 to run a full job hunt, real salary and hiring-manager data, and a plugin to any agent. Fully open source, 0% markup. The next post in his thread links the repo, https://github.com/superdesigndev/treg.

Introducing Jev @typesafeai powered jobs-api Try it at https://treg.to/jobs-api - One API to any jobs data - $0.03 to run a full job hunt - Real salary & Hiring…

Poster from Annatar.md (@AnnatarXBT)

Guides

Annatar.md @AnnatarXBT

EN

@AnnatarXBT posts about a 14-page paper on building a company brain: a cheap decision layer in front of an expensive reasoning model. The architecture is pulled from the TypeSafe and Anthropic docs. Opus 5.5 thinks, Jev decides, and the code holds the branch. You ask a typed question and get a typed answer. Jev returns it with a calibrated probability in 0.44s for $0.00035. A 0.999 goes straight into the if statement, and about 99% of turns end right there. The same gate, placed next to Fable 5.2, is broken down in the article below. The post quotes an earlier @AnnatarXBT post, which calls itself a full harness guide for Claude Fable 5.2 and Jev engineering: one evening with it and Claude does the work of a small dev team, and no $550/month course is needed. That quoted post links https://x.com/i/article/2120740198405357568.

a 14-page paper on building a company brain: a cheap decision layer in front of an expensive reasoning model the architecture, pulled from the TypeSafe and Anth…

Quoted @AnnatarXBT

started building a benchmark for vision decision models was expecting to trade accuracy for speed... meanwhile... - GPT-6.1 Terra (high): 96/100 correct, 7.98s…

Benchmarks

SkalskiP @skalskip92

EN

@skalskip92 says he started building a benchmark for vision decision models, expecting to trade accuracy for speed. On his numbers, GPT-6.1 Terra (high) was 96/100 correct at 7.98 s average per image, and TypeSafe AI Jev 1.14 was 89/100 correct at 9.64 s average per image. He asks why decision models are so slow.

started building a benchmark for vision decision models was expecting to trade accuracy for speed... meanwhile... - GPT-6.1 Terra (high): 96/100 correct, 7.98s…

Poster from DrQuicheEater (@dumbuffalo)

Apps

DrQuicheEater @dumbuffalo

EN

@dumbuffalo says he integrated the Jev model by @typesafeai as an autonomous tram driver. It is not real-time video. It processes an existing tram drive route and suggests actions.

I integrated Jev model by @typesafeai as autonomous tram driver. Not real time video, it processes existing tram drive route and suggests actions. Very cool.

趣味で書いた論文が通ったので一応 https://www.mdpi.com/2078-2489/17/11/1160 Systematic ReviewのTitle and Abstract screeningにDecision modelのJevを使いましたよ、のやつ。安くて速くてよいですねJev

Benchmarks

松井健太郎 睡眠・精神医学 (Kentaro Matsui) @matsuikentaro1

JA

@matsuikentaro1 says a paper he wrote as a hobby was accepted. He used the decision model Jev for title and abstract screening in a systematic review, and writes that Jev is cheap, fast, and good. He links https://www.mdpi.com/2078-2489/17/11/1160.

趣味で書いた論文が通ったので一応 https://www.mdpi.com/2078-2489/17/11/1160 Systematic ReviewのTitle and Abstract screeningにDecision modelのJevを使いましたよ、のやつ。安くて速くてよいですねJev

We worked with @typesafeai to test Jev in the field. Our Maintenance App lets apartment maintenance techs photograph appliance labels instead of typing in model…

Benchmarks

Elise Labs @elise_ai

EN

@elise_ai says they worked with @typesafeai to test Jev in the field. Their Maintenance App lets apartment maintenance techs photograph appliance labels instead of typing in model and serial numbers. If a photo is unreadable, they need to catch it before the technician walks away. They tested Jev, Sol and Opus on 140 labeled photos to see how quickly and accurately they could flag photos that needed a retake. In the follow-up, Jev caught 32 of 35 unreadable photos and incorrectly flagged two readable ones. Its median check took 94 ms, about 30× faster than Sol, at roughly 2% of the estimated cost. They write that these are training-set results. Next they are testing on new photos and measuring the full workflow from taking a photo to receiving a retake prompt.

We worked with @typesafeai to test Jev in the field. Our Maintenance App lets apartment maintenance techs photograph appliance labels instead of typing in model…

Poster from ananmay (@ananmaysharan)

Apps

ananmay @ananmaysharan

EN

@ananmaysharan says he is sketching interfaces with pen and paper, using a decision model, Jev, in real time.

Sketching interfaces with pen and paper using a decision model (jev) in real-time

Poster from Dhaiwat (@dhaiwat)

Apps

Dhaiwat @dhaiwat

EN

@dhaiwat says to be careful on the internet now, because someone is going to post “LUCIA DEATH SCENE GTA 6” on YouTube on day 1. He built a Chrome extension that filters his Twitter and YouTube feed in realtime for spoilers, at https://veil.dhaiwat.com. In the reply under the post he writes that this is built using Jev from @typesafeai.

be careful on the internet now, there's 100% gonna be some idiot posting "LUCIA DEATH SCENE GTA 6" on youtube on day 1 if i were you, i'd use this chrome extens…

You asked for it and we delivered! We just posted the full recording of Fighting Fraud with Jev, our November 4 fireside chat with TypeSafe AI in San Francisco.…

Infrastructure

Sardine @sardine

EN

@sardine says the full recording of Fighting Fraud with Jev is up: the November 4 fireside chat with TypeSafe AI in San Francisco. Fraud and risk leaders spent the evening on why models built for chat wobble on approve-or-decline calls. Erik Gafni, TypeSafe AI’s co-founder and CTO, pinned the problem on RLHF, which trains a model to give people answers they like. His team uses RLCD to train Jev for calibrated calls that stay put when you ask the same question again. Soups Ranjan, Sardine’s co-founder and CEO, showed how that consistency pays off in sanctions screening. They built an AI engine that follows each bank’s SOP and now clears 90% of manual reviews at some of the banks they work with. It has not missed a true match in six months. Yoko Li, a partner at a16z Infra, moderated and pressed Soups on how they catch fraudsters who probe the models. AI shopping agents give themselves away by clicking the dead center of every field, with no natural mouse movement. Soups pointed to tells in their own hiring pipeline: up to 75% of the AI and ML candidates applying to Sardine show fraud signals, down to deepfakes in live video interviews. Erik urged anyone running AI millions of times a day to break the problem into small decisions and give the model as little to do as possible. They made Jev available as a decision node in Sardine Flow so fraud teams can build that way. The recording is at https://www.sardine.ai/events/fighting-fraud-with-jev-on-demand and the recap is at https://www.sardine.ai/blog/fraud-decision-model-stack.

You asked for it and we delivered! We just posted the full recording of Fighting Fraud with Jev, our November 4 fireside chat with TypeSafe AI in San Francisco.…

Poster from Saif (@saifcodes_exe)

Benchmarks

Saif @saifcodes_exe

EN

@saifcodes_exe says two AI models do not write text. They return probabilities. Jev (TypeSafe) versus Gemini Decision 1 (Google), on the same 600 questions: accuracy 86.3% versus 85.8%, a tie; speed 489 versus 621 ms median; cost 28% cheaper on Jev; trust at ≥90% confidence, 93.9% versus 97.6%, and Gemini wins that one.

Two AI models that don't write text. They return probabilities. Jev (TypeSafe) vs Gemini Decision 1 (Google), same 600 questions: Accuracy: 86.3% vs 85.8%, a ti…

Poster from Jordan (@nolansym)

Guides

Jordan @nolansym

EN

@nolansym posts a new AI pattern, Jev Composite Scorecard. Ask atomic score questions, then weight them in code. The demo weights are clarity 40%, completeness 40%, and risk 20%, on a PR description. He built it with @aisdk v7 and the @shadcn registry.

New AI Pattern - Jev Composite Scorecard 🤝 Ask atomic score questions, then weight them in code. ✅ Clarity 40% · completeness 40% · risk 20% ✅ PR description de…

Poster from irony.aditya (@0xironyAditya)

Benchmarks

irony.aditya @0xironyAditya

EN

@0xironyAditya says @liquidai released another model, so he tested d1.2-ultra against the other decision models on two real jobs: grading storm and fire damage on before/after satellite shots, and sorting Serengeti camera-trap bursts into empty or which animal. The setup is 200 picture decisions, the same pictures and the same questions. d1.2-ultra versus Jev 1.14 versus gpt-6 luna versus clef, all through @OpenRouter: exact damage grade 47% / 46% / 42% / 41%; send a team first 70% / 64% / 68% / 69%; animal or empty 99% / 96% / 96% / 98%; species 62% / 63% / 62% / 65%; cost for all 200, $0.018 / $0.034 / $0.034 / $0.083. He writes that accuracy is a near tie at the top, and that d1.2-ultra got there at half the price of Jev and luna and almost 4.7× cheaper than clef. The only catch is that it was the slowest of the four. These cannot replace the person doing the job yet, but they can help sort the pile, double-check calls, and clear out the empty frames.

@liquidai released another banger, so had to test the d1.2-ultra against the other decision models on two real jobs tasks: - grade storm and fire damage on befo…

Poster from Adrián Sáez (@adriaansaeezdev)

Apps

Adrián Sáez @adriaansaeezdev

ES

@adriaansaeezdev says he is tired of politicians lying to his face. He built VerifyLive, a tool that uses jev/clef to detect claims, search for evidence, and check them in real time against neutral sources. You can give it a YouTube video, share a tab's audio, or use your own microphone. The app listens, transcribes, identifies the claims worth checking, and looks for evidence in the press and in official sources, while the debate is still going. In the thread he writes that Verify is not meant to be a truth machine. He does not want another AI that answers with confidence even when it does not know. If it does not find enough evidence, it says so. The model does not have the last word: a deterministic policy limits the verdicts according to the quality and quantity of the sources. He links the repo, https://github.com/adriaansaeez/verifylive, and a live demo, https://verifylive-adriaansaeez-verifylive.apps.nan.builders/. The project was built at the #2 Hackathon of NaN Builders, and he writes that it can be used with a NaN subscription and a JEV API key from @typesafeai until CLEF is available again.

Estoy cansado de que los políticos me mientan en la cara. He creado VerifyLive, una herramienta que utiliza jev/clef para detectar afirmaciones, buscar evidenci…

Jev @typesafeai 열풍이 조금 사그라진 듯하지만, 저는 다양한 실무에 사용 중입니다. 머신러닝이나 스크립트로 해야 했던 판단을 쉽고 빠르게, 게다가 저렴하게 할 수 있다는 점이 정말 큰 장점입니다. 최근 추가된 Jev Box 기능으로 이런 것도 가능합니다. https://br…

Apps

RYAN H. with AIB @AIBVOTE

KO

@AIBVOTE writes in Korean that the Jev @typesafeai wave seems to have cooled a little, but he still uses it across practical work. Being able to make judgments that used to take machine learning or scripts, easily, quickly, and cheaply, is a big advantage. The recently added Jev Box feature makes this kind of thing possible too. He links https://brickfinder.hsu3046.chatgpt.site/, whose title is “Brick Finder with Jev”.

Jev @typesafeai 열풍이 조금 사그라진 듯하지만, 저는 다양한 실무에 사용 중입니다. 머신러닝이나 스크립트로 해야 했던 판단을 쉽고 빠르게, 게다가 저렴하게 할 수 있다는 점이 정말 큰 장점입니다. 최근 추가된 Jev Box 기능으로 이런 것도 가능합니다. https://br…

Pokemonster, my Pokemon card snipping bot got a new helper this afternoon. Jev, a small model from @typesafeai AI. It screens every listing before anything gets…

Apps

Marcos Placona @marcos_placona

EN

@marcos_placona says Pokemonster, his Pokemon card snipping bot, got a new helper this afternoon: Jev, a small model from @typesafeai. It screens every listing before anything gets priced. 168 eBay listings took 3 seconds and cost about $0.013. It flagged a £2.95 “Jolteon ex SIR” as a likely proxy (86%). It also said that same listing was 93% worth pricing. Each question is answered on its own, so he combines them in code. He asks if anyone else is running a cheap model as a filter in front of the expensive one.

Pokemonster, my Pokemon card snipping bot got a new helper this afternoon. Jev, a small model from @typesafeai AI. It screens every listing before anything gets…

Chance says 146 films should be both a slasher and a romantic comedy, and my lab confirms one. This is how I built the map that finds pairs like that, and why t…

Apps

oddstep @oddstepdev

EN

@oddstepdev says chance would put about 146 films in both slasher and romantic comedy among 92,374 films, and the lab confirms one: Who Is Killing the Great Chefs of Europe? (1978). He calls a pair like that a story gap, two common kinds of story that almost no film combines, and the lab finds 116 of them. The first version sorted films by country, because plot embeddings pulled on names and places. To label the story instead, he compiled 207 questions and used TypeSafe AI's jev-1.14 to answer all 207 for every film from its title, year, country and plot: about 19.1 million answers in about 54 minutes, for $18.54. A film gets a label when the probability of yes clears a cutoff, 0.65 for most labels and 0.8 or 0.9 for a few broad ones. Twelve questions worded as general truths made the model say yes every time. Rewording them to ask about the film in front of it cost another $3.75 and brought Con artist from 100% of films to 7.9%. The article's credits say typesafe/jev-1.14, through OpenRouter, answered the 207 label questions for every film. gpt-6.1-sol judged the pairs, and Claude Opus 5.5 wrote the 116 pitches.

Chance says 146 films should be both a slasher and a romantic comedy, and my lab confirms one. This is how I built the map that finds pairs like that, and why t…

I hope y’all don’t mind nerding out with me for a minute... I figured out how to geolocate AND authenticate 98% of the content my system pulls in for http://Thi…

Apps

Micah Berkley - TheAIMogul @MicahBerkley

EN

@MicahBerkley says he can geolocate and authenticate 98% of the content his system pulls in for ThisIsViceCity.com. GTA 6 fan art is everywhere, and half of what looks like a real Rockstar billboard in Miami was made on a laptop. If it goes on his map, it has to be real. His AI crawlers grab a photo or a video still, with no GPS and no caption. @ClaudeDevs Haiku reads the clues: street signs, murals, storefronts. His agent pulls 15 to 20 Google Street View frames of the candidate block and hands them all to JEV at once. He writes that JEV Omni takes up to 16MB of images in one shot, and thanks @Cloudflare. It narrows down the exact spot in milliseconds, then pins it on his custom map. If the scene matches the street, it is real and it gets an address. If it matches nowhere, it is fan art. He says he did not have to ask any human whether it was real. One photo in, an exact address out.

I hope y’all don’t mind nerding out with me for a minute... I figured out how to geolocate AND authenticate 98% of the content my system pulls in for http://Thi…

FWIW, here is the reviewer and you can see it reviewing itself in the PRs. https://github.com/joeharris76/theseus/pull/98 Jev sizes each PR (trivial…max), picks…

Agents

Joe Harris @joeharris76

EN

@joeharris76 points at the reviewer and says you can see it reviewing itself in the PRs, at https://github.com/joeharris76/theseus/pull/98. Jev sizes each PR, from trivial to max, picks routine versus deep capability and the review template (standard, deep, or gate), and decides whether to retry, escalate, or stop when a reviewer fails.

FWIW, here is the reviewer and you can see it reviewing itself in the PRs. https://github.com/joeharris76/theseus/pull/98 Jev sizes each PR (trivial…max), picks…

Poster from Tanishq (tk) (@tanishqk)

Agents

Tanishq (tk) @tanishqk

EN

@tanishqk asks whether Jev demos are still cool. He writes that more coding agents are going toward a “one manager” thread, naming @cursor_ai projects and @AmpCode spaces, and that the old problem is the classic iMessage problem: simultaneous unrelated conversations and back-and-forths in the same thread. He built a demo using Jev that automatically determines whether it should fork the current chat, switch to a previous chat, or stay in the current one.

are jev demos still cool? More and more coding agents are going towards the "one manager" thread approach: @cursor_ai projects, @AmpCode spaces, etc. but the pr…

Here's how I use Jev from @typesafeai for my product: I need to structure Romanian laws for agentic search. But legal text is unpredictable; deterministic parsi…

Apps

Alexandru Ică @vg_head

EN

@vg_head says he uses Jev from @typesafeai to structure Romanian laws for agentic search. Legal text is unpredictable, and deterministic parsing only goes so far. He detects mentions of laws himself. The hard part is how they are mentioned: repealed, introduced, suspended, and so on. Some of that can be decided deterministically, and a lot cannot. Romania's legislation dates back to 1989 and contains over 1.5 million pages, so he needs a cost-efficient way to do all of it and to keep ingesting new publications. With Jev it cost a bit over $1 to structure the whole legal corpus, in less than a day. He writes that the quality was excellent, including in Romanian, which he did not expect.

Here's how I use Jev from @typesafeai for my product: I need to structure Romanian laws for agentic search. But legal text is unpredictable; deterministic parsi…

Poster from Erik Kokalj (@erik_kokalj)

Benchmarks

Erik Kokalj @erik_kokalj

EN

@erik_kokalj tested the new Gemini Decisions API on a recycling line and says it is more accurate, and more expensive, than the other multimodal System 1 alternatives he names, Jev and OAI. Accuracy on this clip: Gemini Decision 1 76%, OAI Decisions API (GPT-6 Luna) 53%, Jev 1.14 72%. Cost per 1,000 images / median speed: Gemini $0.13 / ~340 ms, OAI $0.08 / ~270 ms, Jev $0.08 / ~130 ms. He says @typesafeai’s Jev still has the best performance per price.

Gemini just released Decisions API, and wanted to test it on recycling line ♻️ It's more accurate (and more expensive) compared to other multimodal "System 1" a…

Here is the summary of overall performance between Google Decision 1 and TypeSafe Jev across our four decision primitives. Across all 40 benchmark images, Jev s…

Benchmarks

Flying Higher AI @flyinghigherai

EN

@flyinghigherai posts a summary of overall performance between Google Decision 1 and TypeSafe Jev across four decision primitives. Across all 40 benchmark images, Jev showed higher accuracy and cleaner calibration in multi-class root cause analysis and desiccation severity scoring. The post quotes an earlier post by the same author. That quote reports the Boolean/Noul primitive: Decision 1 at 82.5% accuracy, with 6 false negatives and 1 false positive; Jev at 85.0% accuracy, with 4 false negatives and 2 false positives.

Here is the summary of overall performance between Google Decision 1 and TypeSafe Jev across our four decision primitives. Across all 40 benchmark images, Jev s…

Quoted @flyinghigherai

Jev 之类的 decision model 真的非常好用,我把输出 JSON 判断类的 LLM 全都换成了 Jev / 开源 Jev clone。AI Studio 这个 decisions API 我刚试了感觉挺不错的。 AI startup 的价值是什么?把一些学术界的观点转化成爆款产品,引领行业巨头跟进。今天,…

Notes

Bojie Li @bojie_li

ZH

@bojie_li says decision models such as Jev are genuinely useful: he replaced every LLM he used for JSON judgments with Jev or an open-source Jev clone, and a first try of the AI Studio Decisions API felt pretty good. He asks what an AI startup is worth, and answers that it turns academic ideas into hit products that lead the giants to follow. A SOTA model is outdated in six months, so counting on one product to keep laying golden eggs is unrealistic. TypeSafe may not make much money from Jev, but leading the industry to see why decision models matter is enough.

Jev 之类的 decision model 真的非常好用,我把输出 JSON 判断类的 LLM 全都换成了 Jev / 开源 Jev clone。AI Studio 这个 decisions API 我刚试了感觉挺不错的。 AI startup 的价值是什么?把一些学术界的观点转化成爆款产品,引领行业巨头跟进。今天,…

Quoted @sundarpichai

Unofficial developer resource. Not affiliated with TypeSafe AI. These posts belong to their authors. jevnotes.com does not host the videos.