TypeSafe AI exits stealth with $40M seed and System One model Jev

TypeSafe AI said it raised about $40 million in seed funding led by DCVC and opened early access to Jev, its first “System One” model built for fast, structured, type-safe decisions that software can call directly — with calibrated confidence scores — instead of open-ended chat text. Co-founders include Diogo Almeida (former OpenAI researcher, RLHF/ChatGPT co-inventor), Erik Gafni, and Sasha Sheng.

FINANCE desk — same-day primary seed for a frontier lab explicitly building non-chat models for production software automation, with a named OpenAI-alumni founder and DCVC lead.

Founders, as the Business Wire has it: Diogo Almeida, co-founder and CEO — a former OpenAI researcher the company calls a co-inventor of RLHF and ChatGPT methods — with Erik Gafni and Sasha Sheng. The same about-box says TypeSafe was founded in 2024 and is headquartered in San Francisco. RLHF means reinforcement learning from human feedback — the method that trained language models to follow instructions and chat. Extra team-page titles stay in Sources.

First public model, still company: Jev, named for Jevons Paradox — the idea that making something cheaper can increase how much of it people use. TypeSafe calls Jev a “System One” model, a Kahneman-inspired label for fast, structured judgment rather than a chat assistant. The company says Jev is built for type-safe structured outputs software can consume directly, with calibrated confidence and probability scores on every answer, and that it gives up free-form string generation. Type-safe here means outputs must match a predefined schema so software will not crash on a weird string.

Speed and price, company claims — not independent Bad Signal benchmarks: the company blog says end-to-end latency is roughly 70 to 500 milliseconds, input pricing is about $0.042 per million tokens, and output tokens are free. A millisecond is a thousandth of a second. The same blog claims large speed and cost gains versus frontier chat models on System One-shaped tasks.

Training and sampling, still company: TypeSafe names the training method Reinforcement Learning for Calibrated Decisions, or RLCD. Calibrated here means a higher confidence score is supposed to mean the model is actually right more often. The company says a parallel sampler generates all outputs in one query rather than token-by-token, the usual chat-model method of writing one word at a time. File RLCD and that parallel-sampler picture as TypeSafe’s. How the sampler is wired stays in Sources.

Early access, company: TypeSafe opened a waitlist at typesafe.ai and says it is bringing developers off that list as it can. The company pitches “machine-native, composable AI” as a software primitive — a building block other programs can call — for classification, routing, scoring, extraction, and guardrails. Guardrails here means checks that score or block unsafe prompts and outputs.

Named voices on the company pages: Diogo Almeida, co-founder and CEO; and James Hardiman, general partner at DCVC. Their quotes are color only and stay in Sources. This is not investment advice.

Plain English for the rest of the card: seed = early venture money. System One here = fast structured judgment, the company’s Kahneman-inspired label, not a chat assistant. type-safe = outputs must match a predefined schema so software will not crash on a weird string. RLCD = Reinforcement Learning for Calibrated Decisions. calibrated confidence = a probability that is supposed to match how often the model is actually right. Jevons Paradox = the idea that making something cheaper can increase how much of it people use.