Ask the right question
Choose between rule, Choice, Noul, Score, or generation, with explicit context, question, and options.
The problem
The model answers with confidence. And it errs with the same confidence.
Generating text for a task that only needed a choice is expensive and hard to check. The problems below appear when the decision has no format, policy, or measure.
- You ask for a classification and get a paragraph that your code can't read.
- High confidence becomes automatic authorization, and the wrong answer slips through without review.
- The calculation looks only at the token price and forgets fallback, human review, and deployment.
- Hand-tuned examples become proof of quality, with no independent human reference.
- The API key ends up in the browser and each part of the system calls the model in its own way.
After the training
By the end of the training, you can:
Use confidence without blind trust
Define thresholds, abstention, and human review, and send doubtful cases to a person.
Do the whole calculation
Add up input, fallback, review, and deployment, and choose a pilot that pays for itself.
Apply it to real cases
Tickets, transcripts, clauses, evidence, model and agent routing, and element selection in the browser.
Measure quality for real
Build a human reference, compare alternatives, and freeze thresholds before looking at the result.
Integrate without losing control
Credentials on the server, a gateway with a daily spending cap, cache, cost logging, and failures that return to human review.
Lab + 3 course tracks + 2 projects · 36 lessons
From the case in the browser to the decision in your code.
Modules 1 to 8 are conceptual. Modules 9 to 12 use JSON, terminal, and Python. The lab comes first so you can see the decision at work, and the two projects at the end take Jev into your system.
- 1
Lab
JEV Lab
20 cases with context, questions, criteria, and a simulated answer: edit, import a request, and export the result, with no key and no API.
- 2
Understand
Track 1 · Understand
Where Jev comes in, the three types of question, confidence and error, and the cost calculation that picks the first pilot.
12 lessonsOpen track 1 - 3
Apply
Track 2 · Apply
Support and transcripts, documents and evidence, model and agent routing, browser and sensitive applications.
12 lessonsOpen track 2 - 4
Build and evaluate
Track 3 · Build and evaluate
Integrate without mixing responsibilities, measure quality, operate and observe, and the final project with the adoption decision.
12 lessonsOpen track 3 - 5
Gateway
jev-gw
A single door for all queries from your system: daily spending cap, cache for identical requests, cost and latency per call.
- 6
Comparisonnew
Decisions × Jev
What is the same and what changes between Jev and the OpenAI Decisions API, with the Python bridge that converts questions from one format to the other.
Who it's for
Made for you if
- You have a queue of messages, tickets, or documents that someone classifies by hand.
- You want to use AI in a process and need to decide with numbers whether automating is worth it.
- You integrate models into a system and want to separate deciding, generating, and executing.
- You route tasks between models, agents, or skills and want explicit criteria.
Not for you yet if
- You need generated text, such as a customer reply or a summary: Jev only chooses between options.
- You want to find where the business loses money before automating: start with the Raio-X de Margem area.
Included in the training
You don't leave with theory alone.
The lab, the guide, the gateway, and the packages are open INEMA projects on GitHub, with tests and reports. The videos explain the comparison with OpenAI.
Projects and kits
Videos
Why this training
What changes compared to testing a model in the chat.
Exercise with answer key
12 labs with fictional, original data. You solve them first and check the answers afterwards.
Candor about what was measured
The 20 cases are simulated and are meant for learning. The page separates simulation, rules baseline, and tested real integration.
Packages by area
17 packages with context, criteria, request, and fixture: support, sales, documents, meetings, agents, and others.
Start offline
The lab and the packages run without a key. The real query to OpenRouter or TypeSafe comes in only when you turn on --live.
Questions
Before you start.
Do I need to know how to code?
Not for modules 1 to 8, which are conceptual. Modules 9 to 12 use JSON, terminal, and Python 3.10 or higher.
In what order do I do them?
Lab, track 1, track 2, track 3, and then jev-gw and the comparison with OpenAI. If you already integrate models, you can go from the lab straight to track 3.
Will I spend API credit?
Not in the lab or in the packages without --live, which use simulated answers. Real mode requires a key on the backend and may consume credits.
How much content is there?
There are 36 lessons in 12 modules, 3 per module, plus 12 labs and a final project with a rubric.
What is Jev?
TypeSafe's structured decisions model. The lab and the course are INEMA's independent educational project about it.
Your next AI decision can come with probability, policy, and proof.
Start with the lab: pick one of the 20 cases, change the question, and see what happens. Then track 1 explains why.