TypeSafe AI

TypeSafe AI develops Jev, a model designed to return typed decisions that software can use directly. CEO Diogo Almeida announced Jev on September 15, 2026, after two years in stealth. The company calls this model family System One Models; Jev is its first public model, released in early access.

Its API accepts application state and typed questions rather than asking the model to generate free-form text. Choice selects among options, Score evaluates a rubric and Noul estimates whether a statement is true. Structured values and probabilities let application code use the result; confidence helps determine when to act or defer. TypeSafe calls its training method Reinforcement Learning for Calibrated Decisions (RLCD). Decision accuracy still needs evaluation on the application’s own tasks. The company also publishes an LLM-backed comparison adapter, separate from the Jev model.

Founded in 2024, TypeSafe’s founders are CEO Diogo Almeida, CTO Erik Gafni and COO Sasha Sheng. Almeida previously helped develop reinforcement learning with human feedback at OpenAI; Gafni brings experience building production AI systems, while Sheng worked in research engineering at Meta/FAIR. Almeida’s technical writing examines how models learn: his analysis of distillation distinguishes learning from a teacher’s token probabilities, copying its output tokens and imitating final answers without access to hidden reasoning.

The team works in San Francisco and has investor backing. TypeSafe joined the 2024 AWS Generative AI Accelerator, whose 80 participants were selected from more than 4,700 applications. The program offered each participant up to $1 million in AWS credits, alongside technical mentorship and support bringing products to market.