Jev System One Models: Glean, UiPath and Workday Bake Fast Model Calls Straight into Their Code

A familiar pain: an LLM in product code behaves like a lottery. Diogo Almeida, co-founder and CEO of TypeSafe AI, put it this way at the AI Engineer Code Summit: 95% of the time you get Einstein, and the other 5% you get Mr. Bean. And if you can't tell one from the other, there is nothing to automate. His answer to this problem is called System One.

What System One models are

These are non-generative models. They don't write coherent text and don't reason step by step. They return a structured result: one option from a list, a score or a probability. That is why they are, as a rule, noticeably faster and cheaper than reasoning LLMs, which the conference called System Two.

A simple example. Previously you asked an LLM to "assess whether this document is relevant to the query," got back a paragraph of reasoning, and parsed it with regular expressions. Now you call a model that immediately returns a number or a label. You can put it in a loop, in an if condition or in a sort, and it will behave predictably.

Jev is the product that, according to Latent Space, started this wave. It launched in September, and the podcast with Almeida became the publication's most popular episode of the year. Almeida's main thesis for engineers: code is the language of automation, so the model needs to be embedded in code. Latent Space founder swyx put it this way: sprinkle System One calls generously throughout your application instead of "gluing strings together and hoping for the best."

What is already being built

Glean uses System One alongside LLMs and other classifiers. Chinmay Goyal noted Jev's main advantage: zero-shot flexibility. When the decision taxonomy changes, the models don't need to be retrained again and again. Glean users don't care what's under the hood, and the company deliberately hides the difference.

UiPath showed a "tournament" pattern. Jev picks the most relevant document fragments from small groups, the winners advance to the next round, and so on until the final. The whole process took less than a second, even though there were thousands of Jev calls. The UiPath engineers' takeaway: on its own the model may not seem very smart, but in the right arrangement of such blocks you get a smarter system.

Workday engineer Scott Roberts showed a prototype in which the interface starts filling in a vacation request while the user is still typing. A fast model call here doesn't just speed up the form, it changes the behavior of the software itself: from predicting what you want to say to predicting what you want to do.

Limitations

This is a conference report, not a benchmark. The material contains no accuracy figures, and "95% Einstein" is a joke from the stage, not a measurement. Jev has existed for less than two months, so it is too early to talk about long-term operation. Moreover, Almeida directly compared the current moment to the early internet, when experiments are valuable simply because they are possible. That is inspiring, but it is not a product justification. Reasoning LLMs aren't going anywhere either: Glean itself uses them together with System One.

What to do in practice

1. Go through your code and find the places where an LLM is used as a classifier, ranker or router. These are the first candidates for replacement with a fast model.
2. Try the tournament scheme for ranking: many cheap calls on small groups instead of one expensive call on the whole list.
3. Measure accuracy, latency and cost on your own data, without relying on conference talks. For reliable automation, what matters is not average accuracy but how predictable the result is.
4. Think about where an instant response changes the interface. Autofilling a form as the user types is exactly such a case.

The main point of this story is not a specific product but a shift in architecture: not every decision in an application needs to be handed to a large generative model.