TypeSafe AI, a San Francisco startup founded by an early ChatGPT contributor, left stealth this week with $40M in seed funding and a contrarian thesis: the industry tuned its models for the wrong audience.
DCVC wrote the seed cheque. Co-founder and chief executive Diogo Almeida spent his OpenAI years on reinforcement learning from human feedback and the research that became ChatGPT, and he built the company in 2024 with Erik Gafni and Sasha Sheng.
Almeida’s claim is that conversational models make poor components. A model that shifts its approach between requests, or invents an answer, breaks the software calling it. “People can’t be the only consumers of intelligence,” he said. “Most intelligence should eventually live inside software, running quietly in the background.”
The first model, Jev, is named after Jevons Paradox – the observation that greater efficiency can increase total consumption rather than reduce it. TypeSafe positions it as a software primitive rather than a chatbot: applications call it whenever they need semantic judgment, and it returns calibrated confidence scores so a program knows when to act on its own and when to defer.
TypeSafe says Jev matches frontier quality at under 100 milliseconds of latency, and claims up to 100 times the speed and lower cost, returning hundreds of outputs from a single prompt.
Those are vendor figures. The real test arrives when developers place Jev inside production systems where a confident wrong answer has consequences.
