10 Cases REAIS com JEV

Author: NoCode StartUp · 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: 10 Cases REAIS com JEV · Channel: NoCode StartUp · 12:27 · Watch on YouTube

Topics: AI agents, Enterprise AI applications, Task and workflow automation, Data, databases and embeddings

Summary

The video introduces JEV as a fast, low-cost AI model designed for classification rather than open-ended writing or conversation. It explains how JEV can label inputs, assign scores, and return structured results that other software or models can act on. Practical examples include customer support triage, resume screening, community comment classification, and fraud review. The video then surveys reported projects involving article and flight classification, games, support tickets, marketing analysis, SEO, and LinkedIn outreach.

Key points

JEV is presented as a classification model: it chooses among options or returns a score instead of writing essays, emails, or code.
The video emphasizes speed and cost savings, while noting that uncertain classifications can be escalated to a human.
Its three described output types are a yes/no-style confidence value, a choice from a list, and a score or level.
Customer service teams could use it to identify repeat customers, detect complaints, and estimate customer irritation.
Resume screening can use classifications and weighted criteria to prioritize candidates.
Community comments can be categorized as praise, questions, criticism, or spam, with relevant questions routed for an answer.
A fraud-review example shows how low confidence can trigger human review rather than an automatic decision.
The video points viewers to GitHub, Twitter, and a site collecting JEV case studies as sources for further examples.
Reported applications include classifying articles and flights, making game decisions, processing support tickets, analyzing marketing creatives, and scoring LinkedIn messages.

Chapters

Start	Chapter	Summary
00:00	What JEV is and how classification works	The video introduces JEV’s reported speed and cost advantages, explains that it classifies rather than generates open-ended responses, and outlines its output types.
03:05	Practical classification examples	Examples show how JEV could triage customer messages, screen resumes, categorize community comments, and flag uncertain orders for human review.
07:16	Where to find JEV case studies	The presenter points to GitHub, Twitter, and a case-study collection site before beginning examples from real projects.
08:18	Projects using JEV	The video reviews projects for real-time text classification, article and flight selection, game decisions, support-ticket handling, marketing analysis, SEO, and LinkedIn outreach.
12:27	Closing	The presenter sums up the video’s aim of explaining JEV and invites viewers to share ideas for what to create or publish.

Quotes

While the PT chat writes an essay, Jev goes there and marks an X on the answer sheet. — 01:02

And it agreed with the expensive model’s judgment 90%of the time. — 02:04

Man, I really don’t know if we need to block it or if we can proceed, so let’s pass this on to a human agent — 06:45

So, basically, you can put any type of structure, from customer service to sales, to have a real-time classification there. — 08:50

If you already have any ideas about what to publish or create, comment below so I know. — 12:27

Concepts and entities

Jev, Vercel, Twitter, LLM

Full transcript (original language)

00:00 On September 15th, a San Francisco company launched an air model that refuses to write like chatpt or Cloud, as we’re used to. In 24 hours, it became the fastest adopted model in Gate Versel history, with double the adoption of the GPT 5.6 family and six times faster than Fable 5.1. The numbers the company published are out of this world. In their own evaluations , this model was up to 193 times faster and 444 times cheaper than the

00:31 top-of-the-line model you probably use today. And each response comes out in just a few milliseconds, and the demand was so high in the first few days that the company itself couldn’t even handle VIPI requests. If you try to access their website now, you can’t register. Behind it is a former OpenI researcher who helped build the training methods that became chatpt. And the company just raised $ 40 million in a single round just to bring this to the world. His name is Jeff. Let’s understand what it is, why it’s so powerful, and, most importantly, practical use cases. To

01:02 make it clear, Jev basically classifies with mastery. You give it a situation, which could be, for example, a message from a client, a resume, or just a comment, a post, or a line in your spreadsheet. And you can deliver all of that in a list of possible answers that you yourself give it. It returns which one is correct. Basically, a classification, it’s like a multiple-choice test. While the PT chat writes an essay, Jev goes there and marks an X on the answer sheet. It doesn’t answer open-ended questions, doesn’t write emails to your client, doesn’t generate code, doesn’t talk to

01:33 anyone. If you go over there and say hi and ask how he is, there is simply nothing that he can possibly do. It’s much simpler, but behind that lies something very powerful: speed and scalability. This classification process may seem super simple, but it has an absurd impact on morning costs. For example, the lab called Good Start Labs took 6,000 real-world quality validations. These are the ones companies run to check if a chatbot’s response was good or not. And they did a direct comparison. Running 1 million checks with the Danropic model would

02:04 yield 33,000. Running the same 1 million with Jeff, the count is 160. And it agreed with the expensive model’s judgment 90%of the time. 33,000 versus 60 for the same work. Look at the difference. But let’s go, I’ve brought some R $ 15 to make the power of Jeff clear to you in practice, especially for those who aren’t developers and don’t have technical experience. So let’s understand with some practical examples, okay? But first of all, you have to understand an important definition, which are Jeff’s three points. Basically, they are three

02:34 questions to help with this classification. The first one is the “ no,” okay? It’s the “no” parameter. Basically, it will tell you if that’s true or not. It will return a number from zero to one that A itself will be able to classify. If it’s close to one, it will be yes. If it’s close to zero, it will be no. And close to the middle, it’s unsure about that classification. We also have the choice, which is basically any of these options. We are going to pass a list of some classification options over to our model for processing. You’ll write the list and it will mark which grade or

03:05 point is correct on that list. And the third point is a score, which is basically a level. So we’ll describe the levels in Portuguese and it will return our position. Basically, it will be a classification, a structure to pass on, to run some code or to run some other model. Maybe it’s a little confusing, so let’s go to some practical examples, okay? So, in this example here, of a WhatsApp customer service triage, a message might arrive from a client saying: “Good morning, this is the third time I’ve written here, my access dropped again and I have a meeting with the client in an hour.” So, they’re trying to access

03:36 your product, your community, right? Jeff isn’t going to talk to the client or respond, okay? He’s basically going to structure a classification. So, basically, he would ask the question, right? If that client is already a regular customer. Here, based on the phrase, he gave it 0.91. So yes, he’s almost 100%sure that that client is a regular customer. He has another question here, right, if they’re complaining about something. So here he gave it yes, they’re complaining about something. And here, basically a score to see how irritated they seem to be, right? Here it gave 0.68, so 1.7 out of 2. Basically, it’s a score that he’ll

04:08 pass on, saying that the client, right? He decided that the old client, yes, he’s complaining, a high level of irritation, and these three responses are very confident. He goes straight to service; since the classification is already done, it goes directly to the owner at the top of the queue. So this is a very simple example, but just so you understand that it’s a structure that will classify and structure to give quick answers. But these answers are very fast and you can process a large volume of data, okay? In addition , I’ve included another example here, which is a resume screening. So, the message arrives here: 8 years of

04:39 digital marketing, led a team of four people, last company, lives in Fortaleza, has already set up automations. So you are able to structure a Jeff AI to classify and provide better answers to effectively select the best candidates for the position, is that correct? For example, here, how many years in the area? So you can define the classifications. Here it came to 0.96, so more than five , so he’s closer to one. Here he has experience with the tool the job requires. It came to 0.89, yes. Has he led a team before? 0.93? Yes. Does he live near the job? So he answers and

05:10 then we can move forward. And from this result, we can also define weights, right? In the case of the resume, I can say here that leadership is worth 40%, that daily time is worth 30%, that the tool is worth 20%. So it will always give a numerical result there, see? In this case, it’s 9.1 out of 10 on the interview list. So the priority goes to the front and he will move ahead of everyone to do this interview. Here’s another example of classification. We also see comments in the community. So man, this video saved my project. I just have one question, can this be used without a credit card? So let’s

05:42 suppose here in our model, right, it will classify this as praise, question, criticism, or spam. So we’ll have parameters for that, right? So here it classified that the question is more on the question side, does it deserve a real answer? Yes. So the code will decide which one goes into a shortlist of answers for the day. So basically this model will be able to take and do the first stage of structuring, classification, and then run code or run other processes or even other LLMs, okay? It may seem simple, but at scale it will save a lot and it’s super fast.

06:13 Guys, if you’re enjoying this content, comment here what you think of Jeff, if these cases make sense to you or in your day-to-day work? Comment here, I want to know. The last case I put here is an order that looks like fraud. So let’s suppose we have to analyze whether that order could be fraudulent or not, right? So we have the order here, such value, first purchase by the customer, delivery in Manaus, payment in São Paulo, three attempts until the card is approved. What will Jeff answer ? Basically, he’ll classify it. So, we understand here the delivery address matches the billing address, so 0.04. No, the purchase pattern is atypical?

06:45 Yes, the purchase pattern is atypical. How risky does it seem? So it seems like 0.49. So here’s the point. The risk score is high, but the confidence is 0.49. So the model is divided. In this case, the classification is like saying: “Man, I really don’t know if we need to block it or if we can proceed, so let’s pass this on to a human agent, for example, to gather more information .” So, basically, this classification structure is what allows you to do all of this here in a very quick manner and with a significantly large volume. Okay ? And now that you’ve understood some

07:16 simple cases and some areas, let’s see real things, okay? What people are doing so you can see the power of this. So here, I’ve separated three places for you to find cases and research more . Here we have a repository package on GateHub, where we can see several case studies, okay? Later I’ll leave all the materials below so you can access them, but here are several interesting cases for you to understand what’s happening. And besides that, one of the best sources of information in the field is always, for example, Twitter, right? So

07:47 new content always appears here, people post about what’s happening. So, the first place in the world where things happen in AI is on Twitter, okay? So here you’ll find several case studies to help you understand what’s going on. I found a really cool website that basically brings together all the JV case studies, okay? So I’ll also leave that in the materials description later . And here it summarizes several tweets that have the case studies of the people, what they are doing with Jeev, right? So we can understand the power of this, to see if it’s really just hype or not, right? Sometimes it’s just

08:18 noise, so here we can see it organized by category, which is really cool, okay ? So we have all this structure in these three origins. And speaking of the cases, I’ve separated some more general ones about classification, others more related to games, because there’s all this about decisions, and another about projects in specific areas, okay? So here, the first example I’ve separated for you is from Steve; he basically made a real-time classifier using Jeff. So, to help you understand a little more practically what it is, right? He puts a phrase here, and that phrase is already

08:50 classified here in relation to how it’s structured, genre, tone. So, there’s this whole real-time classification structure. So, basically, you can put any type of structure, from customer service to sales, to have a real-time classification there. It’s very important to see this in real-time sales lead analysis sometimes. We also see the support team doing several projects of this type. So, something cool for us to do. And here’s another example from Rassan, I even translated it into Portuguese here, but he used Jeff to classify 1018 articles. So if

09:21 we were to use a common AI, right, a GPT, a cloud, man, it would take a long time, depending on the complexity of the articles, and it could be quite expensive. So here we see that he spent 0.0808 C and it was very fast. So, basically, he talks about the workflow, right? He summarized each article, sent the title, and he managed to do all this structural classification. And here he left all the links, if you want to know more, and he left here all this structure of this classification that basically did all this analysis of the articles. Another very interesting example here from Gregor, okay? Also translated into Portuguese, he

09:52 basically made a classifier to be able to decide here, choose, find, right, the best cheapest flights. So, basically a whole structure that classifies whether that flight is what he wants, whether it’s cheap or not. And it manages to be super fast, in a few seconds he already chooses the flight and can already book it. So he can do all of that, the whole structure , for 0.0039 cents. So it’s very fast, and we could do this in large volumes, on a large scale. Here I’ve included another example from Marcos. He basically created a Tetris game. So, JS can make the choice here regarding how

10:23 it will be played, how it will be structured. So, he can do that very quickly too. And in these game examples , we also have Aball, who also made a game that classifies the choice of the video game, right? So, basically, he can understand which classification he has to go to the side, up, to the right . And there are several strategies here to go to the side. Another example here is from Justin, okay? He basically made a Tesla car, which is a game where he can analyze if it’s going forward, to the side, if he can, if there are cars nearby or not. So he talks a little

10:55 here about how he created the Tesla part, the self-driving with Jev. So basically, another example of what can be done. Speaking of daily case studies , this profile here created a JV structure for the support area. They opened 40 tickets, nine support tickets , and so on. They even put something more visual running here, which has this selection process. They say it’s very fast and only cost $ 0.016, so it could be applied to other areas. And in the marketing section, MEF posted that

11:25 Jeff is insane. Basically, in 40 seconds, he analyzed 724 creatives. So we can see all this analysis running here; of course, it’s a bit more visual , but he can do this structured classification super fast so you can then make strategic decisions or some automation flow. And another example here for the marketing area, we see a post from Ocara, but he says he managed to create the entire SEO structure, get cited by Chattar Content, and find

11:56 clients on Reddit. So he also created this whole structure, right? It’s a bit more visual, but basically a structure that runs at a very high volume and is super fast, allowing him to handle all these content and marketing elements. Another example from Roman: he basically created a whole structure to analyze what would be the best type of message to interact with his potential clients on LinkedIn, and he was able to score everything within this structure, analyzing and classifying which message to send. That’s it, folks. So, the goal of this video was to give you a little

12:27 more information about what JEV is and how you can use some case studies to broaden your understanding of this new technology that’s just starting in the market. If you already have any ideas about what to publish or create, comment below so I know.