Company and platform guide · TypeSafe System One and Jev introduction
TypeSafe AI is building models for software, not chat
Its System One API gives developers small AI judgments they can compose like programming primitives.
TypeSafe's core idea is simple: keep deterministic work and side effects in code, then call Jev only where the application needs semantic judgment over unstructured state.
- Flagship model
- Jev
- Question types
- Choice, Score, Noul
- SDKs
- Python and JavaScript
- API
- HTTP /v1/systemone
How TypeSafe works
Developers send Jev the current state of an application and a map of narrow questions. The possible answers are described up front, so the response is structured without asking a text model to imitate a schema.
Questions in the same request are evaluated independently and in parallel. The application decides how to combine them, what thresholds are acceptable, and when uncertainty requires review.
What TypeSafe provides
- A hosted System One endpoint with jev-latest as the default model alias.
- Python and JavaScript SDKs with typed question and response objects.
- Patterns for intent routing, confidence gates, composite scoring, extraction, reranking, and verification.
- Cookbooks that show how to keep workflow logic in code rather than inside a prompt.
Start building
Begin with one decision your application already makes poorly with rules or expensively with an LLM. Supply only the state needed for that decision, ask one narrow question, and retain the full probability output.
Test the result on representative data before choosing thresholds. Add more independent questions in the same call only when each answer remains useful and interpretable on its own.