Jev is TypeSafe AI's System One frontier model: unstructured state in, typed probabilistic decisions out. Instead of generating text, Jev returns Choice, Score, and Noul answers with calibrated probabilities your code can act on. Parallel sampling delivers ~70-500ms responses, about 20-200x faster and 40-400x cheaper than comparable LLM workflows, with output tokens free. Now available to everyone at console.typesafe.ai with no waitlist.
I run 10+ products solo and read the support inboxes myself. After a press release for one of them, the contact address got buried in pitches: webinar invites, free whitepapers, agencies wanting to run my ads. Gmail filters didn't hold. Senders change every time, and "thank you for your inquiry" shows up in both auto-replies and real human replies.
The worst ones borrow a platform. Someone shares a pitch as a Google Doc, so the email comes from Google. Or they sign you up for their Zoom webinar, so the invite comes from Zoom. You can't send Google or Zoom to spam.
Jev is what finally fixed it. It doesn't write anything, it just tells you how likely each yes/no answer is, and you decide where the line goes. Per email I ask five yes/no questions in one call: did a person write this, should we reply, are they selling us something, do they want a meeting, are they clearly asking us to stop. A few rules around it:
Ask twice, act only if both agree. Otherwise I look myself.
Hi, I'm Rikin, maker of dotpals (a small open-source pal that shows what your AI coding agents did). Most people seem to use Jev for lead-gen and booking calls. I'm using it for something different: as a judge for test results. My coding agent says "All tests pass. Done!" far too often. dotpals reads the test runner's own output first, and when the output and the exit code disagree, it sends Jev a short, redacted snippet and asks one question: "did these tests pass?"
It answers in about 0.2 seconds with a probability. In the first day it caught a real false "pass" (a run where the tests never actually ran) at 92% certainty, and it rated our own 148-test suite as passed, 98% sure.
I run a large job marketplace in Japan, and Jev has been in production on it for about 10 days. Not as a reviewer, it's making real calls in live jobs.
The first place was the company list we keep in house. We check each company name on it against the operator name on their website. Code settles most of them, and only the leftovers go to Jev, with just the two names (about 200 tokens). On the first run it settled 571 of 996 companies, and 50 hand checks found 0 mistakes. Since today it also reads replies in our inbox: is it a person, and does it need an answer.
I have a side project that trades on prediction markets. It pulls in news, social posts and on-chain data, then uses an LLM to estimate the odds of each outcome.
The weak spot is the middle step. Most of what comes in is noise, and asking an LLM to read everything and write back a probability is slow and hard to trust.
