Use case · Triage and routing
Jev for triage and routing
Picking a route, a label or a next model — classification where a wrong answer is recoverable.
47 items, from every source we cover All use cases
- GitHub
GitHub · project
anishfn/shapeshift
The clearest design pattern in the batch, and the author states it in one line: Jev decides, code computes. Fourteen typed questions in a single call decide what the text box should become, while dates, amounts and units are parsed by deterministic code.
-
Hacker News · thread
Benchmarking Jev: what a decision model can (and can't) do in an agent harness
The most useful dev.to piece on this model so far: about 22,500 calls across ten public datasets, with the code and raw results in a companion repository. Half the value is the negative results — model-difficulty routing at 51% (no signal at all), trajectory failure attribution at 0.560 AUROC (random), and non-English retrieval well below English. The other half is two engineering lessons from iterating: decompose a fuzzy judgment into orthogonal questions and compose them in code, and let a choice compete before a noul verifies.
- YouTube
YouTube · video
Building a Harness with Jev
A harness build rather than a model review: how to put the decision call inside an agent loop, what to send it, and what to do with an answer that arrives as a probability. Two hundred thousand views, and the most practical of the three.
-
Hacker News · thread
Decision models like Jev don't beat LLM-as-a-judge or traditional classifiers
Red Hat's AI Safety team benchmarked nine guardrail candidates across four methodologies — CPU-scale pre-trained classifiers, LLM-as-a-judge and decision models — to ask whether Jev-style models earn their place for risk detection. It is the first enterprise-scale comparison we have seen on this question, and its answer leans toward the cheap old thing rather than either vendor's story.
- GitHub
GitHub · project
devagrawal09/jev-review
Read the README for the pipeline rather than the dashboard: a Null risk matrix, then Choice and Score file profiles, evidence selection, mechanism classification, severity, and a conditional Choice that routes which reviewer sees it. Orchestration stays in code and Jev only answers the bounded judgements — which is the pattern to copy.
-
X · post
Everyone has been impressed by TypeSafe AI’s Jev — I wanted to see for myself how good it is …
A product engineer testing the model on the part of the job everything else depends on: whether it understands the text before it decides anything about it. The post is a sequence of measured probes rather than a verdict, which is why it is on the wall.
-
Hacker News · thread
Fixing Jev Decision Model AI Agent Overrides: A Flutter/Node.js Blueprint
A debugging write-up: an agent's override logic fighting the model's decisions, and the pattern that fixed it across two runtimes. Narrow, concrete, and the kind of thing that only exists once somebody has actually shipped.
-
Hacker News · thread
How I stopped misclassifying jobs with Jev!
The most concrete adoption story in this batch: a job-board operator replaces regex filters with a Jev judgment layer and names the miscategorisations it fixed — Java is not JavaScript, remote is not hybrid. First-hand, and it says what the old filter got wrong rather than only what the new one gets right.
-
Hacker News · thread
Human, agents, code, judge: adding Jev without replacing peer review
A team adding the model to a review process without letting it take over the review, and describing where the human stayed in the loop. Most integration posts skip that half; this one is about it.
-
Hacker News · thread
I Built a Fraud Agent using Jev That Can Trace a Transaction Across 590K+ Transactions
A fraud-investigation agent where Jev decides whether to look and the graph decides what is true, with deterministic rules and human approval keeping the last word. The architecture is the useful part: it shows where a decision model belongs in a pipeline that has to be auditable, and where it must not.
- X
X · video
I built a Sales Copilot using Jev → Listens to sales call live → Tells you what to say next →…
A live sales copilot that listens to a call and decides what to say next — a real-time loop where a wrong answer is expensive and a slow one is useless. The proof is a demo video; there is no evaluation in the post.
-
Hacker News · thread
I got Jev to zero mistakes. I'm still using Flash-Lite.
An eleven-minute field report whose conclusion is in the title: the author got Jev to zero mistakes and still ships Gemini Flash-Lite. This is the useful shape of a negative result — not 'the decision model is bad', but 'here is the job where the cheap general model was already enough'.
- X
X · video
I think I just cooked something 🔥 jev(): a PostgreSQL extension that searches your whole dat…
A Postgres extension that judges rows in natural language with no index and no embeddings: one function inside a WHERE clause, a hundred and twenty-nine rows in about a second. The demo is small and the shape is not — a filter that can read what a column means.
-
Hacker News · thread
It took me 18 hours to add Jev
Eighteen months to build the pipeline and eighteen hours to add the decision model to it, retold without a victory lap. The ratio is the point: this model's integration cost is small enough to be a weekend, and that is why so many of these rows exist.
- YouTube
YouTube · video
Jev - The Ultimate Classification Model?
Reads Jev as a classifier, which is the closest everyday description of a typed-decision model. Worth watching for that framing alone.
- X
X · video
Jev + Kimi K3 for fraud detection! TLDR: Jev classified 100 emails in 1.42 seconds, then I ro…
Fraud triage in two stages: Jev labels a hundred emails in 1.42 seconds, then the uncertain ones go to a large model — ninety-six of a hundred right for about seven cents. The accuracy is self-reported; the routing pattern is the part worth copying, and it is what the whole site argues for.
-
X · post
Jev could become the control layer AI agents have been missing. Instead of spending 5–20 seco…
Argues that Jev belongs as the control layer rather than inside the model call. The seconds figure in the post is the author's framing of agent latency, not something we measured.
- X
X · video
Jev Engineering is what turns an agent stack into an actual control system and moves the expe…
A video version of the same argument: Jev as the thing that turns a stack into a control system. Watch it for the vocabulary, which is spreading faster than the model is.
- X
X · image
Jev Engineering: how to build the fastest AI Agent Brain in 10 Steps (Full-Setup)
A ten-step article about putting Jev where an LLM was doing the cheap work inside an agent loop — the same case our builds column keeps finding, written as a walkthrough with the setup included. The numbers in it are the author's, not ours; 750 likes and 31 replies at the time we read it.
- X
X · image
Jev is cool not because it re-invented classification, but because it makes ARBITRARY classif…
The most precise framing we found — not a new classifier, but arbitrary classification made cheap enough to put inside a loop. Read this before the vendor's own claims.
- X
X · video
JEV is INSANE. We gave it 700 high-intent leads and personalised outreach messages. In 40 sec…
Seven hundred leads and the outreach message written for each one, scored in forty seconds for nine cents, with mismatches between message and lead flagged. Judging a written sentence against an intent is the errand; the accuracy claim is the author's and there is no holdout in the post.
-
X · post
Jev is the "Internet" moment for the AI industry It tells your agents and LLMs what to do nex…
The strongest version of the hype, posted as text. Read it for how the analogy is argued; the post itself offers no number, and none of the claims in it are ours.
- X
X · video
Jev solved local harness/model routing I use a combination of Claude Code, Codex and Opencode…
A developer using a decision model to route between Claude Code, Codex and opencode: the practical version of the routing argument made elsewhere in this column.
-
Hacker News · thread
Jev vs Claude for OpenPoke: 5.8x Faster Email Screening in a Small Test
A small, honest test — one task, a handful of emails, a stated speed ratio — from somebody building an assistant. Small tests like this are how adoption actually happens, and the post does not dress it up as a benchmark.
-
X · post
jev will make agents 10x faster and cheaper, here's how: 1/ model routing: pick the right mod…
A thread on model routing as the first place a decision model pays for itself. The 10x in the opening line is the author's claim, not a number we checked.
-
Hacker News · thread
Jev-Driven SRE Diagnosis: What Worked and What Failed
A research group handed Jev a view of a cluster and let its fast judgments choose where to investigate and which evidence supported a diagnosis, then published all 105 diagnoses including the ones that went wrong. Negative results with the failure cases printed are rare in this topic, and worth more than another demo.
- YouTube
YouTube · video
JEV: How It Works and What You Can Build
The only walkthrough in this column that publishes its own chapter list, and the order is the one we would suggest: routing at 2:08, then what Jev is, then the three answer types, and cost and speed only after that. Twenty minutes, and the first four are the part a developer needs.
-
Hacker News · thread
LLMs Generate. Jev Decides. Software Should Know the Difference
An integration write-up rather than a benchmark, and the framing is the reason it is here: the author's complaint is that we took a machine built to generate open-ended sequences and asked it politely not to be. He put Jev into a real Java invoice-processing application through The Pipeline Framework, found the old prompt already contained the list of answers, and reports that some of the work being sent to an LLM was never an LLM problem. Declared AI-assisted on dev.to.
-
Hacker News · thread
ML vs LLM for Classification: Jev, Decision Models, and the Triage Bot I Overbuilt
A first-hand account of overbuilding a triage bot and then cutting it back once the author understood what the model actually returns. Useful precisely because the first version was wrong and the post says so.
-
GitHub · project
realZachi/pg-jev
A PostgreSQL extension that answers plain-language questions about your tables: Jev does the choosing, the database does the work. It puts the decision model inside the database rather than beside it, which is the same architectural bet Cloudflare made by putting Jev behind env.AI.run.
-
X · post
Seems to me that the best application for @typesafeai System One models will be: *the AI rout…
The argument that the model's home is the router rather than the task: every automation stack already has a controller, and that controller is where the cheap decisions pile up. It is a post about architecture, and it is the clearest statement of the case in this batch.
-
X · post
Stanford CS PhD here 👋. I rigorously benchmark Jev/Typesafe on my consumer AI app serving 2.…
A Stanford PhD running a hiring product with 2.5 million monthly users benchmarked the model on the task his company actually pays for: scoring how well a résumé matches a job. Production numbers on a real workload are rare in this topic, and he shows the method.
- GitHub
GitHub · project
superagents-lab/jev-search
Search where Jev chooses the sources, the time range and the terms, and then ranks what comes back — links and snippets with visible relevance scores, no generated answers. There is a live instance, so this is the one row in the column you can check in a browser without cloning anything.
-
Hacker News · thread
Testing TypeSafe Jev, Mistral and Gemini for local event validation
A team with early access tested all three on their own task: validating local event data. The comparison is narrow and the write-up says so, which is more useful than another sweep of public benchmarks.
- X
X · video
this model is actually insane at email classification i tested it on 1500 of my own emails to…
One person running fifteen hundred of his own emails through the model and reporting that it works well. A personal inbox is a fair place to start and also the reason the post carries no accuracy figure: the only labels in it are his own.
- GitHub
GitHub · project
thruwire/foreman
Puts the decision model above the coding agent instead of inside it: a Codex or OpenCode worker does the engineering, and Foreman independently asks whether the ticket is finished, the tests exist and the requirements were met. The separation of the worker from the thing that checks the worker is the pattern most of our field notes end up describing.
- GitHub
GitHub · project
TianyuCodings/NanoJev
A small replica of the mechanics: parallel decisions, dynamic routing. Useful as the readable version of the architecture if you are not going to work through the paper.
-
Hacker News · thread
TypeSafe’s JEV Model: Is It Really 193x Faster and 444x Cheaper?
The independent check on the two numbers everyone quotes. They ran 100 support tickets through four question types against three frontier models — 400 decisions each — and measured 4x to 7x faster and 31x to 65x cheaper, not the 193x and 444x on the launch page. Their explanation is the useful part: a multiple is a property of the comparison, and the vendor compared against slow multi-step workflows while they compared against mid-tier single calls. They also reproduce the vendor's own side, a 474 ms median against the published 70-500 ms.
-
Hacker News · thread
Valid Schema, Wrong Content: Using Jev to Guard a Google ADK Agent
A three-part write-up of putting the model in front of an agent as a guard, and discovering the failure mode that matters: the answer had the right shape and the wrong content. The schema was never the hard part.
- X
X · image
We benchmarked fx auto mode (safety) classifier with @typesafeai's Jev. tl;dr: ~5-18x faster …
A benchmark post with a concrete result: roughly 5-18x faster on a safety classifier. Their numbers, their harness; it is listed because they published the setup.
-
Hacker News · thread
Why Most Type-Safe Validation Fails in Production (And How JEV Fixes It)
The practical version of the argument: validation rules written as code rot, and a decision model is one way to move the judgement without moving the structure. The title overpromises and the body is a decent field report from a production system.
- GitHub
GitHub · project
YaoApp/yao
A self-hosted agent workspace with eight thousand stars that added Jev as one of its decision models, announced in the README rather than in a launch post. Useful as evidence of the pattern reaching infrastructure other people already run.
- X
X · video
Yeah Jev by @typesafeai is very cool. It classified 500 emails in seconds. And it costed 3.5 …
Five hundred emails classified in seconds for a few cents, which is the claim; the number is the author's and we have not run it. The interesting part is the shape — bulk triage where per-item cost decides whether it is viable at all.
-
Hacker News · thread
Your certificate inventory is exact. Your tickets are not.
Certificate renewals mapped to the tickets that track them, with the model deciding which ticket belongs to which certificate. Unglamorous operations work, stated plainly, which is what most adoption of this model will look like.
- GitHub
GitHub · project
zilliztech/memsearch
The largest company-backed project in this batch: a semantic memory layer for coding agents from the team behind Milvus, with Jev as an opt-in reranker. Worth a row because the integration is one bullet in a long feature list, which is how adoption looks when it stops being news.
-
X · post
一些关于 Jev (@typesafeai) 的 notes 研究了一天 Jev,带来的新鲜感迅速回落,这好像就是一个更快的通用分类器/决策器,LLM 完全可以做到。 而且因为不了解模型…
A sceptical Chinese-language note: after a day of study the author reads Jev as a faster general classifier that an LLM could equally do. The counterweight this column needs.
- X
X · video
拿 Jev 做搜索重排,我先泼一盆冷水:单独用,它没打赢向量检索 TypeSafe 的 Jev 这阵子很火,一堆项目拿它做重排。我们在 Agent Skills Hub 的 33,047…
A negative result worth keeping: on their reranking test, Jev alone did not beat vector search. The only post in this column reporting that something did not work.