TypeSafe AI

San Francisco AI lab serving Jev, a model that returns calibrated decisions instead of text.

TypeSafe AI is a San Francisco research lab building System One models, AI trained to make fast, typed decisions for software instead of writing text. CEO Diogo Almeida co-authored OpenAI's InstructGPT paper, and the lab trains with its own method, Reinforcement Learning for Calibrated Decisions (RLCD), so the probabilities its models return are calibrated against real outcomes. The flagship, Jev, answers a batch of questions about your input in one call: true-or-false checks, a pick from labels you define, and scores on your own rubric. On Opper it runs on TypeSafe's first-party System One API with a global scope rather than a pinned region, and existing TypeSafe clients switch over by changing their base URL. Pick this route for classification, intent routing, moderation and LLM-as-a-judge scoring, where your code needs a number to branch on rather than a paragraph to parse.