~/til/jev-classification-model

Jev: Classification Model

Today I learned about Jev, a model from TypeSafe AI built to make typed, structured decisions inside software, instead of generating text.

What is Jev?

Jev is TypeSafe’s first System One Model, and the name comes from Kahneman’s “System 1” thinking: fast, intuitive, gut-check judgments, as opposed to the slow, deliberate hallucination-filled garbage we get sometimes from general-purpose LLMs.

Not only is it fast, but also very cheap. And the output tokens are free.

We must pass Jev a state (some context: a support ticket, a document, a log line, anything), and a set of typed questions about that state. In return we get typed structured answers with calibrated probabilities and a confidence score. No text is generated, just a structured response that we can decode and use without hoping the response matches our desired schema.

There are three question primitives:

  • Choice - pick one option from a defined list (e.g. which team should handle this ticket).
  • Score - rate the state against an ordered rubric (e.g. how frustrated is this customer, from calm to furious).
  • Noul - a yes/no question that returns the probability the answer is “yes” (e.g. is this urgent?).

How to use it

A request is basically “here’s some state, here are my questions”:

curl -X POST https://api.typesafe.ai/v1/systemone \
  -H "Authorization: Bearer $TYPESAFE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "state": "Hi, I'\''ve been trying to connect my Stripe account for 3 days and the integration keeps failing. I'\''m losing sales. Please help ASAP.",
    "model": "jev-latest",
    "questions": {
      "department": {
        "type": "choice",
        "instructions": "Which team should handle this",
        "criteria": {
          "billing": "Payment or subscription issues",
          "technical": "Bugs or integration problems",
          "sales": "Pricing or account questions"
        }
      },
      "is_urgent": {
        "type": "noul",
        "instructions": "The message conveys urgency or time-sensitivity"
      }
    }
  }'

The response comes back fully typed, with a probability per option and a confidence score:

{
  "answers": {
    "department": {
      "choice": "technical",
      "confidence": 0.78,
      "probabilities": { "technical": 0.85, "sales": 0.0, "billing": 0.15 }
    },
    "is_urgent": { "noul": 1.0 }
  }
}

Applications

  • Routing: which team/queue should this support ticket, email, or lead go to.
  • Content moderation / guardrails: flag toxic, off-topic, or policy-violating content before it reaches a human or another model.
  • Scoring: sentiment, urgency, lead quality, risk/fraud score — anything that slots into a threshold-based decision.
  • Extraction and tagging: pull structured fields out of unstructured text at scale (map-reduce over large datasets).
  • Verifying LLM output: use a cheap classifier to judge whether a bigger model’s answer is safe/relevant/on-topic before you show it to a user — a judge in front of a judge.
  • “Smart if-statements”: anywhere your code currently has a brittle hand-written rule that’s really trying to approximate a fuzzy judgment, a small classifier can replace it and get recalibrated as data changes, instead of rewritten by hand.

The common thread: when correctness just needs to be “good enough and fast,” and the output has to be consumed by code rather than read by a human, you don’t need a frontier chat model — you need a classifier that returns a typed answer and a confidence you can threshold on.

References

Dario Camargo

Software Engineer
Go, Java, Python

Currently at Mercado Livre
Proximity / Delivery Platforms

Location: São Paulo - Brazil - UTC-3

Languages:
- Portuguese (Native)
- English (Fluent)
- Spanish (B2)

Dario standing

Projects

Today I Learned