TypeSafe SDK - Jev, the first System One model from TypeSafe

Oh wow! It seems like it’s a race to the moon with all these model releases these days!

Also, congratulations on releasing this SDK. So much time and effort went into the README, which is very commendable! I’m sure the actual code is even more straightforward.

Adding this to my watchlist. Thanks for the work!

Thanks. Yes, this is all over AI social media. Huge game changer for anyone working with agentic workflows.

Quick follow-up to my earlier typesafe_sdk 0.1 post: 0.2.0 is out:

TypeSafe SDK v0.2.0

This release turns the initial TypeSafe/System One client into a much richer Elixir-native semantic SDK: strict noul / choice / score constructors, prepared questions, evaluate / evaluate_many / streaming evaluation, caller-key preservation, request-relative response validation, richer probability/Score helpers, supervised concurrency, a real transport-level testing API, telemetry, schema tooling, and a labeled evaluation workflow. The original system_one API remains compatible.

Perhaps unsurprisingly, given the noise and relevance: three other independent Elixir TypeSafe clients appeared almost immediately around the launch. I reviewed all three and leveraged the best features from all three for this release:

Matt Neel – https://github.com/mattneel/typesafe: response ergonomics, ranked distributions/margins, expected vs. modal Scores, validation paths, schema export, telemetry and testing ideas.

hfiguera – https://github.com/hfiguera/typesafe_ai: request-relative response validation, retry/replay discipline, privacy-conscious telemetry, runtime-boundary thinking, and model-vs-application-policy evaluation.

TypeSend – https://github.com/typesend/typesafe_ai: semantic question DSL, caller-key preservation, prepared questions, enriched Score semantics, batch evaluation and application testing patterns.

Thanks to these authors for their contributions to the ecosystem. Acknowledgement is in the README.md.

Any success stories with it? It does look interesting.

Yeah, but most things are not in written for the BEAM, of course.

Fast browser agents are hot. Video game agents. LLM guardrails, code review, rdbms natural language filters. There are a few “awesome” lists on github.

I’m working on a boring Elixir coding agent sandbox project using it. It’s a good fit for the tech – fast/performant typed semantic judgment for security boundaries.

Thanks for the shout-out, @gtcode ! Hello all…I would love any feedback on my latest version and docs, just published:

https://typesafe-api.hexdocs.pm/0.1.0-alpha.3/readme.html

4 new libraries on the same day for the same thing - it’s a good sign Very keen to play with them all - on the waitlist.

Dannote has released a fifth version, and, as expected, it includes some really neat features:

https://github.com/dannote/jev

All five clients are MIT-licensed, which is a reasonable concession to the reality of modern AI development. I’m borrowing some of Dannote’s features in the next release.

ReqLLM shipped support for Jev today as well, v1.24: req_llm | Hex

Here’s a gist I put together to walk through it: Evaluate issues with Jev using ReqLLM and a plain Elixir GenServer · GitHub

It’s a different beast, so opted to expose it via a new evaluate/4 method

In case it’s of interest to anyone, re Jev, saw this on Devtalk: Laya - the OS version of Jev, created a year ago - AI In The News - Devtalk

And then in September 2026, a well-funded frontier lab called TypeSafe AI (founded by Diogo Almeida, a co-inventor of ChatGPT at OpenAI) launched Jev. They proposed the exact same non-autoregressive decision concept as if it was a brand-new scientific breakthrough. Except they launched without technical papers, without open weights, and with zero open training datasets.

Source:

Haven’t read it yet - but so Jev is not the ‘first’ as per their marketing?

Thanks. There are a number of alternative open weight alternatives. I’ll look into streamlining a safetensors/elixir/nx mini ecosystem around them, then turn typesafe_sdk into a generic system one harness.

Relevant discussion on hn: