What Is Jev? Explained in 10 Mins!

Author: Pavan Belagatti · Watch on YouTube

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: What Is Jev? Explained in 10 Mins! · Channel: Pavan Belagatti · 09:58 · Watch on YouTube

Topics: AI agents, AI developer tools, AI models

Summary

The video presents Jev as a model designed for bounded decisions rather than generating long-form answers. It describes use cases such as routing support tickets, scoring search results, and estimating yes-or-no probabilities, with structured outputs that applications can use directly. A playground demonstration shows a choice, a relevance score, and a notification decision, including probabilities and confidence values. The speaker notes that Jev can make mistakes and that generative LLMs can also provide structured output when configured.

Key points

Jev is presented as a model for making defined decisions from a situation and a specific question.
Users define possible choices, and the model returns a decision with probabilities.
It can also score results or estimate the probability of a yes-or-no answer.
The video contrasts bounded judgments with generative models that may return explanatory paragraphs.
A support ticket about duplicate invoice charges is routed to the billing team in the demonstration.
A search result for waterproof trail running shoes receives a relevance score of 1.69 and confidence of 0.54.
A calendar reminder example returns a probability of 0.83 for waiting until quiet hours end.
The speaker says Jev can make mistakes and frames reported speed and cost gains as Type Safe’s own evaluation results.

Chapters

Start	Chapter	Summary
00:00	What Jev is designed to do	The speaker introduces Jev as a model for decisions, choices, scores, and probabilities that software can use.
01:33	Why use Jev instead of a generative model?	The video explains how bounded decisions can avoid unnecessary paragraphs and compares Jev’s purpose with generative LLMs.
03:40	Playground and ticket-routing example	The speaker opens the playground and demonstrates choosing a team for a customer support request.
05:48	Scoring search relevance	A search query and product page are scored, with the output showing a relevance score and confidence.
08:21	Estimating a notification decision	The final example uses a probability question to decide whether a calendar reminder should wait until quiet hours end.

Quotes

Give Jev a situation and a specific question and it returns a decision your software can use. — 01:03

When my software needs a decision, do I need a generated paragraph? — 02:06

My application can use the result directly. — 03:09

it will exactly give you what what is required what your application needs — 07:50

see how how the model responds. — 09:58

Concepts and entities

Jev, TypeSafe, LLM, gemini.dev/playground

Full transcript (original language)

00:00 Hey developers, so in this video we’re going to see the AI model that is making rounds and making a lot of buzz in the industry today, that is Jev. So, we will see what is Jev and how it is different from the other models. And also at the end I have a simple simple playground to show you how Jev basically works. So, let’s get started. So, most of us know that AI models as tools that generate something like an answer, an email, or a piece of code. But Jev is built different. Jev has a

00:31 different job. Type safe built it to examine information and make a defined decision, right? It is It is actually built to make decisions, right? So, imagine giving it a customer support message and asking which team should handle this, right? So, you you define the possible teams and Jev returns a choice with probabilities. So, that’s the important thing here. It It will give you the probabilities, right? So, no other model can give that probability thing, right? So, you can also ask it to score something or estimate the

01:03 probability of a yes or no answer also. That’s what Jev does, right? So, that’s the idea I want you to remember. Give Jev a situation and a specific question and it returns a decision your software can use. So, that’s the whole thing. Maybe you won’t understand it as I explain it now, but but when I show you the playground, you will actually understand how Jev model is used for decision making in the applications or in any of the situations, right? Now

01:33 coming to the main thing, right? Why everyone is talking about this particular model Jev, right? So, think about how many small judgments any application makes, right? For example, which team gets this ticket? Is this search result relevant? Should an agent’s next action be checked? So, developers often use a generative LLM model for these questions, which is which is kind of not necessary, right? So, even when the application only needs a category, a score, or even a yes or no probability, so we all use a generative LLM model

02:06 uh where we will get a big paragraph even for a small even for a small query, right? Which is kind of not necessary. So, Jase is interesting because it focuses on those bounded decisions, right? So, Type Safe has also reported substantial speed and cost gains in its own workflow evaluations. Um so, but I would treat those as Type Safe uh reported results rather than a claim that Jase is faster than faster for every AI task. But, the practical question is simpler, right? When my software needs a decision, do I need a

02:38 generated paragraph? So, that’s what Jase does. Like, it will it it will give you a probability of for any of the smallest questions that you have. And that is where Jase basically shines. So, now let’s understand how Jase is different from the other generative LLM models, right? So, Jase versus a generative LLM model. So, let’s use the ticket example. If I ask the generative LLM model, which team should handle it? I might get an explanation in sentences, right? So, that can be useful when a

03:09 person wants to understand the issue. With Jase, I define the allowed teams first. Its answer fits that decision. For example, billing. Along with probabilities for the available options. My application can use the result directly. So, this does not mean Jase cannot make make mistakes. It can choose the wrong team, and an LLM can also return structured output when configured to do so. The distinction I’m showing here is what Jase is built to do, return a

03:40 bounded judgment properly. In a real workflow, an LLM might draft a reply while Jase helps route the ticket. So, that’s the main idea behind uh the model Gem model. So, this is what you see the Gem AI model diagram. So, you can go to gemini.dev/playground. So, if you go there, you can actually try the Gem model, right? So, here is the complete playground.

04:11 As you can see, try the Gem AI model. Choose one question or three different questions about the same context. Try a real scenario. Edit the inputs. Try inspect each type answer, right? So, here explore uh GitHub projects, use cases, subscribe, and all these things, right? Here is Here comes like building a decision, and this is where we are using the Gem AI model. And I’m purposely picking this one question like the choice. We can actually use Gem to

04:44 do decision making or to score or for the probability. But, I’m using it for this particular thing, the choice, right? And the context is just like how I explained. A customer says, “My invoice shows two charges for the same order. Can you refund the extra payment?” Now, we have the choice one, the question or the instruction you can call, “Which team should handle this particular request?” And we have We have the choices here, right? The billing team, the technical team,

05:15 and the general all other requests, right? And uh Billing These comes in the payments, invoices, and refunds. And technical is all about bugs, API errors, and integrations. And general is all other requests, right? Now, if I run this particular decision, I want an answer, right? So, if I run this, let’s see, right? We get a structured output, as you can see here, right? The model response a structured answer. “Which team should handle this request?” The choice is billing, right? So, type

05:48 type is choice. The choice is billing. The probability equals the general is zero, billing is one, and the technically is zero. Confidence is one. So, this is the type of output you’re going to get when you use the Jack model. I can also edit it and I can also do the other other example take the other example. For example, “Rank a search result based result on relevance scale.” The search query is “Lightweight hiking shoes for rainy weather.” Result

06:19 “Product page for waterproof trail running shoes with a size guide with and customer review.” So, that’s the context that our Jack model has. And this is the score part of the question and the instruction. “How relevant is this result to the search query?” And these are the score limits. “Unrelated to the query, somewhat relevant but misses key need, and strong match of all the query.” So, you find the rank this particular decision.

06:50 Here is the model response. Structured answer like how relevant is this result to the search query? The the score is 1.69. The type is score because we we picked the score here. And then the score is 1.69. The legend here which explains like “Unrelated to the query is zero, somewhat relevant but misses misses key need.” That is one. Two, “Strong match for the query.” But the query was

07:20 “Lightweight hiking shoes for rainy weather.” And the result is “Product page for waterproof trail running shoes with size guide and customer review.” So, these are the probabilities that it is getting. So, the answer the score is 1.69, right? So, for this particular search query and also the result, these are the probabilities, right? And the confidence also it is giving 0.54. So, I hope you are getting it. You can use this model in your in your applications when you want the

07:50 probabilities and the confidence, right? So, it won’t give you a particular explanation in the form of how an how a generative LLM model usually gives, like like an explanation or something a paragraph, but it will exactly give you what what is required what your application needs, right? As an output, a structured a proper structured output. So, you can also see like unrelated to the query it is zero, somewhat relevant, but necessity meet is one. And strong match for the query is two.

08:21 The probability of this whole the confidence of this whole thing is 0.54, whereas the score is 1.69. Right? And also I can also go for the probability itself here. Let’s say the context is it is 22 30. A calendar app wants to send a reminder about a team lunch scheduled for next Friday. The user has quiet hours enabled until 7:00. And now the probability the question is should this notification wait until quiet hours end? And then what comes,

08:54 yes or no? So, we are given this input here. False, send now because the reminder is time sensitive. True, is delay until quiet hours end without harming the user. So, if I run this decision, we will again get a structured output here. A structured output. Probability is should this notification wait until quiet hours end? That was the question. I mean, that was the that was the thing that we wanted to answer. And the probability is 0.83. The type is now and yeah, the now

09:27 is 0.83. The probability we just say. So, this is the this is the probability structure that you can use in your AI applications, right? Whenever you are making you want to make a type of decision. So, that’s it and yeah, the same question is repeated. I mean, that’s what we want to answer, right? Should this notification be retained? So, they basically give you the probability answer for your query. So, that’s the whole thing about the Gemini model. I think you can go to this particular gemini.dev playground and to basically play around

09:58 with your own questions also, right? And see how how the model responds. responds, basically. So, that’s it about the Gemini model and thanks for watching.