What Is Jev by TypeSafe AI? Here's How I'm Using It

Today we dig into Jev by TypeSafe AI. I'll give you the quick breakdown first, then show you three ways I'm using Jev right now in hopes that it unlocks ideas and sends you down a path to build your own thing.

I'm not an AI expert. I'm building in public and sharing what actually works. So, without further ado, let's build.

The TLDR on Jev

You give Jev context and defined questions, and you get typed decisions plus probabilities. Your code chooses what happens next.

That's the whole thing. Now, I know it's a bit confusing at first to think, "Oh, is Jev just another ChatGPT that I can go chat with?" Not necessarily. The easiest way to understand it is to put it side by side with your typical LLM.

| | Jev | Typical LLM | |---|---|---| | What it is | A decision model for software | A language model for flexible generation | | Built for | Fast, narrow decisions in code | Open-ended, token-by-token generation | | How it works | You provide context and typed questions, you get parallel decision outputs | You prompt plus context, you get whatever it generates | | What comes back | Choices, scores, or truth estimates. Choice plus score include probabilities and a confidence measure | Its interpretation of the context and question you asked | | Reach for it when | You want to route tickets, score leads, or classify inputs | You want to draft replies, explain ideas, write code, or collaborate |

The keyword is classify inputs. I would define Jev more as a classifier than a large language model. It's trained for calibrated decisions, which means it doesn't operate like a chat model. You have to first set up the prerequisites. You have to decide what metrics you want to measure. As long as you know that, you're golden.

We can define what we measure. That's the beauty of Jev.

And more importantly, the biggest difference is cost and output quality on this kind of work. An LLM will cost significantly more and give you a less desirable output on classification. Jev gives you the output based off of how you configure the pipeline.

The golden loop: Jev classifies, your code routes, the LLM drafts

Jev and your LLM are not competing. They pair well together. You have Jev process the information to find the outputs, and then you have your LLM draft the reply.

Jev classifies. Your code routes. The LLM drafts the reply. That's the golden loop to utilize when utilizing Jev.

Now, let me dig into three quick examples. The purpose here is just to inspire you. I built three unique things, and I want to show them off.

1. A Gmail inbox classifier

This first one is inspired by Riley Brown's video, where he created a classifier for his inbox. Shout out Riley.

I took it one step further and actually linked my Google account. So I tied in the Gmail API alongside my TypeSafe API. In the demo I had already pulled 500 emails, then ran it again, and you can watch it classify in real time into these buckets:

• Brand deal
• Client
• Subscription
• Cold pitch
• Newsletter
• Team
• Other

The second run finished faster than the first. And it was identifying things correctly. A lot of paid collabs in there, a report from Muddy, a cold pitch. Pretty nifty.

There's also an edit classifier screen. This is where I can change the available choices and the question I want Jev to answer about each category. That's the part I'd point you to if you build your own. Make the classifier editable so you can refine it as your inbox changes.

Listen, I'm not going to dig too much more into this one. It's an example meant to inspire you on ways you can utilize Jev.

2. Audience radar for YouTube comments

This is a fun one. I originally built it for our channel, and then I thought, what if I created it for other channels as well?

Here's what it does. You enter a channel and specify how many videos to pull, the max comments per video, the sort order, and how many parallel workers to run. I keep the workers capped at 100 so it operates as quickly as possible. It catalogs the videos that fall within that criteria, checks every comment up to the limit, and then, based off of details and criteria I configured, gives me video ideas based off of that comment section.

I hadn't tested it on Riley Brown's channel yet, so I ran a fresh scan on camera. I selected 50 videos (I doubt he does more than 50 per month), maxed out comments at 100 per video, and sorted newest first. Here's what came back for his channel:

• 10 videos discovered
• 584 comments collected, all classified by Jev
• 75 action signals
• 6 opportunities found

I break the signals down into five buckets: questions, problems, requests, objections, and reactions. Generally speaking, I ignore objections because those are usually trolls or people who disagree with you. There's some value in that, but it's not what I'm looking for. I'm looking at the questions, the problems, and the specific requests.

Then you get into the opportunities. Each one is a video idea with a "why this idea" explanation and evidence, meaning the actual comments that relate to that question. The top ranked idea in the demo was "Where your agent budget actually goes," backed by 45 evidence notes. I can open the original comments and validate it myself.

Under the hood it's a YouTube Data API key plus a TypeSafe API key. It's still a work in progress. I haven't explored unlimited comments yet, and I'm eventually going to add an interest meter per idea. But I've already gotten a bunch of cool video ideas from running it on Clearmud.

And more importantly, I can point it at any competitor's channel and their comments to snipe video ideas from their audience asking them. You see my point?

3. Slate IQ, a daily fantasy sports optimizer

Last but certainly not least. I've been playing fantasy football for over 10 years. It started out as a hobby and now it's turned into a side hustle. I play season-long leagues, but I also play weekly.

The problem is there are a lot of tools scattered with a lot of different features, and I still have to manually go to certain places to find information. So I thought, "What if I build my own web app that utilizes Jev to pull from a data source, classify that information based off of metrics I want to discover, and give me the outputs in my own app?"

That's Slate IQ. Here's what's in it so far:

• Jev intelligence. Collect and evaluate pulls all the players based off of how I've configured it.
• Player intelligence. Select a player and get recent form, performance away, target share, offensive snaps, and weekly targets. It also shows history against specific opponents, like every time a player has faced Arizona, how many targets he got, and how many points he scored across all games.
• Matchups tab. Previous week's snap counts and receiver matchups. Still fine-tuning this one.
• Lineup builder and research lab. The research lab was the first thing I started tracking.

Most of the data is pulled from Pro Football Focus. Normally you'd have to navigate that site manually, go find splits, game logs, Super Bowl performance, and so on. Now Jev processes all of that for me based off of the metrics I want to measure.

Here's the one I think gives me an edge. One metric I track is where a player went to college and what their home state is. I don't have the exact number off the top of my head, but rookies in a starting position who go back to their hometown to play for the first time, I want to say 70% of the time they hit the over and 5x their value. It's an insane metric that not a lot of people think about. Even veterans returning home tend to perform better, because they're buying tickets for friends and family to come see them. Later in their careers I've noticed they just don't care as much anymore. They're happy to see their people at the game.

We're still early in the NFL season, so this gets more valuable as the weeks go on. Over the next month or so I'm going to brain dump my manual research process into Slate IQ, with Jev powering the collection and processing of all that data.

I'm not here to show off the tool. I'm here to show off how I'm utilizing Jev.

Key takeaways

Three use cases, all built for myself. I didn't build these for the video.

• Inbox classifier. Gmail API plus TypeSafe API, classifying my Google Workspace email into brand deals, clients, cold pitches, and more.
• Audience radar. Analyze our comments to come up with video ideas, and analyze competitors' channels to do the same.
• Slate IQ. A passion project that automates my fantasy football research with Jev handling data collection and processing.

One more thing on the stack. I built the front end of all of these web apps with Astra 6, and the outputs are displayed on a web app I built with ChatGPT 6 Ultra inside Codex. Jev sits on the back end classifying and categorizing the inputs. It doesn't matter whether you use Claude Code, Grok, Codex, or Gemini. Pair your preferred LLM or coding agent with Jev, assuming your web app can benefit from its strengths, which are classifying and categorizing inputs rather than having your LLM do it for you.

Because it's faster, it's cheaper, and it doesn't hallucinate, more importantly.

So don't sleep on Jev, but don't think Jev is just another LLM. It's far more than that.

If you'd like to see me make a video around a specific topic, drop a comment on the video. Let me know who you are, what you do, who your target audience is, and what your question is, and I'll add a video to my queue custom tailored just for you.

My name is Marcelo. This is Clearmud, and clarity matters.

Watch the full video: https://www.youtube.com/watch?v=DXR3imONF7Y