TypeSafe AI builds machine-native intelligence infrastructure for developers and enterprises that need reliable AI-powered workflows. The company's System One Models are a rebuilt model stack - new architecture, parallel sampling and a training method it calls Reinforcement Learning for Calibrated Decisions (RLCD) - designed to return typed, probabilistic decisions rather than generated strings. The stated aim is to address overconfidence, mode dropping and unreliable model behaviour in production systems, particularly in latency-sensitive and deeply automated workflows where constant human supervision is impractical.
Jev is the first public model built on the System One Models architecture. Because its outputs are typed and probabilistic, software can use them directly to classify, route, score, extract, branch and verify information, without parsing free text. The work spans reinforcement learning, probabilistic decision models, calibrated machine learning, model architecture and parallel sampling, and targets use cases in software development, enterprise software, AI workflow automation and developer tools.
The company is based in San Francisco and was co-founded by Diogo Almeida, Erik Gafni and Sasha Sheng, a team of researchers and operators with roots in frontier AI research, including InstructGPT and ChatGPT. The founding team spent two years developing the idea in stealth before launch.
