Browse builds

best way to close my day with a webinar on @typesafeai Jev in JarvisCore! i covered the 4 places we are using Jev in our agent runtime: RAG classification and ranking, model routing, subagent routing, and directly in agent decisions this is not the end, i believe there is much more to come in our memory assembly, tool recovery, peer delegation, and human-in-the-loop. Every agent harness, runtime, loop needs some good Jevglue! also check out this https://t.co/nrzsIzpxM8 by @CompleteSkeptic

🤔 Where do #Jev and other "Decision Models" fit into your current agentic or non-agentic pipelines? The short answer: anywhere you might have LLM calls with a #pydantic model or json schema to type-cast your raw data into structured categories, Likert-scale numerical scores, or true/false decisions. I ran three benchmarks with @flyteorg on https://t.co/WczbNQoVLn in the context of three use cases: - Customer support - Code review - Legal contract review In all three cases, using Jev was faster (up to x2.6) and cheaper (at least x10) with no loss in quality compared to using a "System 2"

Event-sourced #ruby app with Jev classifying comments. I had to lower concurrency to the minimum otherwise it's too fast to see 😆

What if Jev do a triage to your GitHub backlog ? 543 open issues. ~26 seconds. $0.17. What came out: → a map of impact against readiness → good first issues for new contributors → duplicate clusters

I turned my Rubric eval platform into a small decision-model lab. Added Jev and started comparing it with chat LLMs on the same structured decisions. First run: Jev hit 100% accuracy, matching GPT-4o-mini, while showing 2.3× lower p50 latency (1.02s vs 2.39s) and 2× lower cost/1K. Then I put Jev into an actual routing pipeline instead of testing it in isolation, query → Jev → RAG / small LLM / big LLM / human, instead of throwing every query at the same model. Interesting part: the handwritten rules router got 100% route accuracy with 0 wrong routes, while Jev got 93.8% with 1 wrong route. Jev

jev is insane i ran 34,186 engineering profiles through it with no boolean search or pre-filtering we gave it a jd and 8 questions, which came out to 273,488 judgements in 2 mins 35 sec

this is what a decision layer looks like when your LLM stops making the cheap calls every task a typed decision a probability distribution, a confidence score, a route, and a signed receipt. ~80ms each. watch the gates do the work: "charged twice, wants the $49 back" refund_full 0.77 AUTO "delete everything + i'm disputing the charge" 0.46 ESCALATE legal-ops, human assigned $0.0011 per decision · ~$0.33 saved vs human triage. jev decides. skills + repos execute. the model only writes when it actually has to

AICodeKing just put TypeSafe's Jev through real support-ticket tests ~8 minutes of Choice, Score, and Noul on messy customer messages he starts with a duplicate-charge ticket Jev picks billing, puts refund probability at 98%, urgency at 11%, and keeps frustration near calm then he removes Other and forces billing / technical support / sales Jev picks sales at confidence 0.31 restricted choices stop invented labels, they don't stop a wrong useful pick Watch it, then read the full guide on building a Jev support-ticket router below

You're missing what Jev is actually capable of, and it's not the model's fault same model, same 20 drafts, two different questions, and only one of them comes back as a decision ask it to "rate the quality of this post" and all 20 come back between 0.4 and 0.6 vague question → every answer in the middle → nothing to sort or filter → a threshold that never fires → "this model is useless" now ask whether the post has at least one fact that would be FALSE about a different company, and the same 20 split into two piles near 0 and near 1 the second answer is a decision your code can act on, the fir

MY MAC NOW READS MY X FEED FOR ME, 1,318 JUDGMENTS IN ONE SCROLL, 203MS EACH, AND IT TOLD ME 48% OF MY TIMELINE IS NOT WORTH READING. The setup is small (Revealing now, might delete later) ⬇️ > Jev running as the decision layer, typed output only > a feed reader that grabs every post while I scroll > 12 typed questions fired at each post in parallel > is_shill, reads_like_ai, claim_needs_source, worth_reading, mute_author > every answer is a bool, a score or a choice, with confidence > verdict per post: keep, reply, skip, mute The output is not an answer, it is a row {"mute_author": 0.80, "con

Made an omarchy app to categorize (g)mail with Jev. It's called `Omajevmail` and a great add-on to `Omamail`. Very simple concept: - link up your mailbox (gmail, hey, outlook, imap,..) - add an openrouter key - set up categories - categorize those emails (using Jev) Emails are then ordered by highest matches per category allowing you to quick archive

We use LLMs for a surprising number of problems that aren't generation problems. Been experimenting with Jev from TypeSafe for tool search and ranking. Wrote about what we tried, what worked, why I think this pattern goes beyond search reranking 👇🏽 https://t.co/tnreET4OEa

THIS ENTIRE TERMINAL WAS BUILT USING JEV + GPT-6 ASTRA. We put it to the test: 100,000 X posts scanned for scams in 18 seconds. JEV + GPT-6 Astra is 750x faster and 800x cheaper. A complete pass with heavy standalone LLMs would burn $620. We processed the whole dataset for 62 cents. Why traditional AI moderation fails: everyone forces heavy reasoning models to write essays when you only need instant classification. Our pipeline bypasses text generation entirely and runs 12 parallel binary checks per post: > brand impersonation signature > wallet drainer trigger > malicious support redirec

I was using Gemma 4 to categorize articles on Bredbox. Just tested using Jev: same results, 127x faster, 1/10th the cost. 🤯

Things I've used AI personas for lately: – synthetic customer messages to test a classifier built with Jev – testing a landing page on 3 different customer segments – roleplaying user interviews before running the real ones Cheap, fast, weirdly accurate. What's the most interesting thing you've done with AI personas?

We rewired our renewals agent's Next Best Action flow with Jev: its 4.5× cheaper, 2× faster, accuracy holding, and a chunk of routing middleware gone. Is this the future of GTM orchestration? https://t.co/zRizyK0EhH

too much ai noise every link feels important every thread feels like "alpha" my research agent was getting overwhelmed too many tokens spent on things that did not matter i needed a gatekeeper so i built a filter using JEV a system one model that does not think or summarize it just decides: yes or no the pizza test: > pepperoni pizza recipe: rejected > technical doc on agent architecture: approved do not let your most expensive model do the job of a simple filter use a fast, cheap decision layer to clean the input then let the researcher do the heavy lifting fewer tokens burned on thi

I built Inbox triage using Jev 🤯 So I point my gmail to jev and it classifies all the emails i have in my inbox and put them in specific category folder. It was built using claude. I also published a small demo video on my youtube. Sharing the repo and youtube link below.

Mitigating risks while using Jev for decisions Jev evaluates content and returns structured answers with probabilities. You supply the content as “State” and define “Questions” with criteria for judging it. I tested jev-1.13.0 on synthetic emails written to influence Jev's decisions that were also written with the intent to exploit an AI email agent, resembling real world use cases. In this threat model, an attacker controls the email body included in the State field, but cannot change my Questions. Simple instructions inside an email could steer its classification, a risk TypeSafe also docu

I have a plugin to evaluate a post to see if it relevant to my interest -> and this is an excellent case for using Jev each post now show: > my original algorithm score, > total number of followers of the author > Jev score yeah, Jev is fast but still not as fast as a deterministic algorithm, for the accuracy, I will need to track more to see how good it is the score. Jev can be a generic/meta classifier, but come to a specific domain, probably a simple algorithm could still win, both speed and accuracy

For the Tano hackathon @corgicafeco @0xkaushik_k and I built an inbox classifier using Jev. A content creator who get 4K + DMs a day. We didn't want ot build another DM bot, but a better way to help manager her DMs so that she doesn't answer the same question 400 times, and only surfaces the ones that need her judgment. All of this, according to her playbook and workflow, so her audience continues to trust her judgment. We were finalist out of 50+ people who turned up on the day. https://t.co/tjHMvQNiHU Thank you @sashacayward @sa64r and the entire Tano team for a great event.

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

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

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

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

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

Built a Claude Code mod using Jev that watches what you're asking and routes the turn to the right model. Quick renames go to Haiku. Real implementation? Opus. Architecture planning? Fable. @claudeai @ClaudeDevs @bcherny @lydiahallie @trq212 https://t.co/DgIgIGpHLq

Built a Ticket Triage Copilot using Jev (@typesafeai) instead of a general LLM for support ticket classification. 🔗 GitHub: https://t.co/zQ9oB7t7Eq

I think Jev really flourishes inside your code more than in the actual product. Right now, I’m using Jev to classify the thinking mode of our AI agent, so it can select a high, medium, or low mode depending on the prompt.

Built an AI dispatcher for Zammad tickets using Jev (TypeSafe AI). New ticket comes in, Jev decides which team owns it with a confidence score attached. Not a chatbot, a structured decision. Bridge is a tiny Ruby/Sinatra service. Open source, MIT. https://t.co/bXdJpGWHuw #jev

Built an SLO-aware inference router using Jev It uses Jev as a typed decision model to select the optimal LLM for each request based on predicted quality, latency, cost, and live backend load Releasing full walkthrough video soon https://t.co/XUjKj9Ryyn

Built jevmetrics — an OTel Collector component using Jev inference to score metric relevance, redundancy, and keep/reduce/drop decisions. Fail-open, with a deterministic policy still in control. https://t.co/1zLhyNjoEp #OpenTelemetry #Observability #Jev #Typesafe

everyone's using jev to route between models. we put it on inbound sales replies the question isn't which model handles this turn, it's whether "sure" means yes. ours used to log that as a rejection and bin the lead [N] real replies through ai gateway. [x]ms p50, whole batch cost [z] papercut on the way: experimental_evaluate doesn't exist in ai v6, evaluation modality is v7 only, and nothing on the gateway model page says so. @vercel worth a line

i just shipped a small repo that will help you understand it perfectly.its a Real-time support triage + response bot using Jev and cerebras https://t.co/l4gFvr0GdW PS if you like it,approve my entry to the hackerhouse. I applied with rewantgoenka87076@gmail.com

exactly!! just shipped a real-time support triage + response bot using Jev and cerebras. https://t.co/l4gFvr0GdW

youre right. For example in a support ticket jev can classify way faster that which agent should get the ticket. I built a Real-time support triage + response bot using Jev and cerebras https://t.co/l4gFvr0GdW

just shipped a Real-time support triage + response bot using Jev and cerebras: https://t.co/l4gFvr0GdW this should help you get started!

I just built a Real-time support triage + response bot using Jev (from @typesafeai) and @cerebras Jev routes the ticket to a specialist agent (general / account / billing / technical) and decides whether a human should take it instead ,Cerebras then drafts the reply using whichever agent Jev picked. Repo: https://t.co/p6y4IQ2VhZ @callmeshuklaji @communidiyi @hackgoofer @dgo_almeida @sydneyrunkle

Jev join the race 🔥 Jev read the top 250 Hacker News stories 250 stories 2000 typed judgments in 16 seconds cost: $0.0095 about AI 32% Negative Tone: 41 ( 16% ) Built with Jev by @typesafeai https://t.co/RY3GiLTks9