AI that answers the question you asked
typeai.dev gives developers small, dependable units of AI judgment — the kind you can put in an if statement, log, and test.
Why we built this
Teams building with large language models keep running into the same problems. Is this prompt an attack? Does this model call need the expensive model? Did this pull request break a rule we agreed on? Is this alert real?
These are judgment calls, and asking a chat model for them means parsing prose, paying for long outputs, and hoping the format holds. We think they deserve a better primitive: a typed question with a probability as the answer.
What we believe
- Code stays in control. The model supplies judgment; your code owns the workflow and the final decision.
- Policies should be readable. Thresholds live in plain code you can review and change, not inside a prompt.
- Uncertainty is information. When the model isn't sure, your system should know — and escalate.
- Collect only what's needed. We store usage metadata to run the service — not the content you send us.
Built on Jev
Our endpoints run on Jev, a "System One" model from TypeSafe trained to make fast, calibrated decisions rather than generate text. We design the questions, combine the answers, and apply the policies — so you get a ready-made tool instead of a prompt to maintain.
Narrow questions
Each tool splits a fuzzy judgment into specific, independently useful questions.
One call
All of a tool's questions are answered together in a single, parallel model request.
Explicit policy
Probabilities become decisions through thresholds you can see in every response.
Let's talk
Questions, partnerships, or feedback — we read every message.