CHOOSEBusiness workflows
Meet Jev: The AI That Sorts the Orders Instead of Writing the Recipe Imagine you love cooking an
FoundOnArdy@FoundOnArdy
SignalA Choice question asks Jev to select from those options.
Pattern↑ Jev AI decision models
Meet Jev: The AI That Sorts the Orders Instead of Writing the Recipe
Imagine you love cooking and open a small restaurant.
At first, you prepare the food, answer questions and check new orders. Then things become busy.
A customer wants to book a table. A supplier says the flour will arrive late. Another customer reports a missing dessert.
You need help, but not a long explanation about every message. You need answers to small questions:
What kind of message is this? How urgent is it? Who should handle it?
That is the kind of work Jev supports. Made by TypeSafe AI, it is an artificial intelligence model that returns structured decisions instead of written explanations.
Let's understand it through our restaurant.
A CHATBOT WRITES. JEV CHOOSES.
A large language model, or LLM, can generate text. You might ask one to explain bread-making or write a recipe.
Imagine a cooking teacher who can describe every step.
Jev has a different job. It evaluates information and answers questions with limited, defined answers. TypeSafe calls this a System One model: its name for this decision-focused approach.
Give your kitchen assistant three possible destinations for an order:
Pasta station. Pizza station. Dessert station.
The assistant chooses where the order belongs. But choosing “dessert station” does not make the dessert.
Jev supplies a decision. The surrounding application decides what to do with it.
THE THREE KINDS OF ANSWERS
1. CHOICE: “Which group does this belong to?”
Suppose your restaurant receives this message:
“Our delivery of flour will arrive tomorrow instead of today.”
You have created four categories: customer booking, supplier message, customer complaint and other.
A Choice question asks Jev to select from those options. This sorting is called classification. You supply the categories and their descriptions in advance.
In our example, “supplier message” would be the intended category.
Think of labelled trays on your kitchen counter. One holds dessert orders. Another holds pizza orders. Classification means deciding which tray receives each ticket.
Clear labels matter. If two trays have almost the same meaning, sorting becomes harder. An “other” option helps when none of the main categories fits.
2. SCORE: “Where does this belong on a scale?”
Now you know the message concerns a supplier. But how urgent is it?
A Score question places information on an ordered scale. You describe each level. Those descriptions form a rubric: the rules for judging the level.
Our restaurant might use:
Low: It can wait until next week.
Medium: Someone needs to handle it today.
High: It affects orders we are preparing now.
A furniture catalogue might belong near the low end. A missing ingredient for tonight's menu might belong near the high end.
Choice identifies the group. Score measures a level.
A score can also fall between levels. It does not always have to be a whole number.
3. NOUL: “How likely is the answer to be yes?”
Now ask:
“Does this message say a delivery will be late?”
A Noul question returns a number between zero and one: the estimated probability that the answer is yes. Probability means how likely something is.
For example, 0.90 means an estimated 90% probability of yes.
It does not mean the delivery is 90% late. It does not measure the size of the problem. It estimates whether the statement is true.
The video transcript spells this “null,” but the official name is Noul. Vercel's integration calls this type Boolean. Here, the question concerns yes or no, while the probability expresses uncertainty.
A CONFIDENT ANSWER CAN STILL BE WRONG
A ticket saying “two chocolate cakes” seems easy to sort.
A ticket saying “the customer wants something sweet but also asked about pizza” is less clear.
For Choice and Score, Jev returns probabilities across the possible answers. It also supplies a confidence value describing how strongly those probabilities favour one answer rather than several.
That confidence number is not simply the probability that the chosen answer is correct.
Think of the difference between:
“This clearly belongs with desserts.”
and:
“Desserts seems most likely, but another station might need to help.”
An application can send uncertain cases to a person. The rules should depend on the consequences of a wrong decision, not confidence alone.
TypeSafe calls its training method RLCD: Reinforcement Learning for Calibrated Decisions. The company's aim is useful decisions with meaningful uncertainty estimates, not just convincing language.
Imagine training an assistant to sort tickets AND recognise unclear ones.
MODEL ROUTING: CHOOSING THE RIGHT COOK
Your kitchen has several cooks. One prepares simple dishes. Another handles complicated meals. A specialist makes difficult cakes.
You would not send every small task to the specialist.
Model routing applies a similar idea to AI. A system decides which model should handle a request. Jev can make that choice from defined options.
A simple sandwich description might go to a smaller model. A difficult question might go to a more capable one.
A prompt is the instruction given to an AI system. The router evaluates the request and passes it to the selected model.
Choosing the cook and cooking the meal are separate jobs.
Jev can help select a writing model without writing the answer itself.
STRUCTURED OUTPUT: AN ORDER TICKET, NOT A PARAGRAPH
Imagine two ways to send an order to the kitchen.
One is a paragraph describing what a customer said. The other is a ticket:
Dish: vegetable pasta
Quantity: two
Table: six
The ticket is easier to use in a repeatable process.
Structured output means information returned in a known shape. JSON, short for JavaScript Object Notation, is one common format for labelled data.
LLMs can produce structured output too. Jev's difference is that constrained decisions are its native output, rather than a special request to a general text generator.
But a correctly filled ticket can still contain the wrong order.
Correct structure and correct judgment are different things.
WHY SPEED AND COST MATTER
Imagine receiving a thousand kitchen tickets.
Writing an explanation about each one creates extra work. Often, you only need a category, a priority or a yes-or-no assessment.
TypeSafe reports large speed and cost advantages in its own tests. It also acknowledges that its biggest reported gains may be near the upper end of real-world results. They are not guarantees for every application.
In his video, Riley Brown shows an application processing 500 emails, assigning categories and assessing importance.
Some messages are flagged as possible scams. A flag is a model judgment, not proof of fraud.
Our restaurant version could sort bookings, identify supplier messages and highlight complaints for review.
The benefit is not perfect decisions. It is making many small decisions practical to check quickly.
CONTEXT: THE INFORMATION ON THE COUNTER
“Tomorrow” might be early enough for one order and too late for another.
An assistant needs context to decide.
In Jev, the information supplied for evaluation is called the state. It can include a message, related records and relevant business rules. The questions tell Jev what to assess.
Picture an order ticket, the delivery schedule and tonight's menu together on the counter.
A context window limits how much information a model can process in a request. It is counted in tokens: small units of text, not necessarily whole words.
The video mentions 64,000 tokens. TypeSafe's documentation specifies 64,000 across the request, with a 32,000-token limit for the state plus the longest question.
The counter is only an analogy. Jev processes data, and it does not automatically know everything about your business.
COULD IT RUN THE WHOLE RESTAURANT?
Not by itself.
An AI agent is software that works towards a goal by taking steps and using tools. Jev can help it choose a next step. Application code controls what happens next.
Deciding that a message belongs with the booking team does not reserve a table. Another part of the application must check availability and make the reservation.
The video also presents browser-control, trading and driving-related demonstrations. They do not establish safe real-world driving, profitable trading or reliable completion of every browser task.
Current Jev inputs are text-based. It does not directly accept images, audio or video. Another tool must first turn those into suitable text or data.
A fast ticket sorter is not a complete restaurant team.
HOW AN APPLICATION CONNECTS TO JEV
An API, or Application Programming Interface, is a defined way for one program to request a service from another.
Think of an order window between two parts of a kitchen.
Your application sends information and questions. Jev returns structured answers. You can connect through TypeSafe's API or Vercel AI Gateway, a service giving applications access to AI models.
An API key identifies an authorised account. Treat it like a private access pass, not something to put in a public tutorial.
A coding assistant can help build the application. Jev then performs its decision-making role inside it.
WHAT TO REMEMBER
Choice: Which category fits?
Score: Where does it belong on a scale?
Noul: How likely is the answer to be yes?
A message arrives. An assistant sorts it, assesses its importance and helps decide where it should go. Unclear cases still need attention.
Jev is not the chef who writes every recipe or cooks every meal. It is a specialised decision-maker that can help the rest of the kitchen work together.
Based on Riley Brown's “JEV: How It Works and What You Can Build”:
youtu.be/o1CogAtWdBk
Technical references: TypeSafe's documentation and launch explanation, plus Vercel's integration announcement:
docs.typesafe.ai
typesafe.ai/blog/introducing-system-one-mo
vercel.com/changelog/typesafe-ai-jev-now-a
Want to learn more about this topic? Search it on foundonardy.com.