TypeSafe AI exits stealth with $40M, DCVC-led seed

TypeSafe AI's pitch rests on a tension the company says is built into today's large language models: the same qualities that make them good at talking to people — flexible phrasing, confident answers — work against them when they are embedded in software that needs consistent, machine-readable outputs. The startup says that mismatch forces applications to rely on human review. Jev is its answer: a model that accepts structured questions and returns typed answers with probabilities and a confidence measure, which a developer can use to decide whether to proceed automatically, seek more information or refer a case to a person. In insurance underwriting, for instance, the company says Jev could estimate the likelihood that a property catches fire and route an ambiguous case to an underwriter.

The company's leadership carries OpenAI credentials: Diogo Almeida worked on reinforcement learning from human feedback, InstructGPT, ChatGPT and GPT-4 at OpenAI before founding TypeSafe in 2024, and he told Forbes that popular generative AI models have been optimized to please humans. DCVC Management Co. LLC led the seed round, which Forbes reported valued TypeSafe at $200 million. TypeSafe says the largest opportunity for Jev is in high-volume processes — classifying service requests, evaluating invoices, triaging security alerts and reviewing the results of AI agents — where a specialized decision model could act as a control layer while a language model handles drafting.

The stakes hinge on whether Jev's confidence scores actually track accuracy on real customer data, something TypeSafe itself says developers must test. The company's speed and cost comparisons — less than 100 milliseconds per result, 39 cents per 1,000 workflows — come from its own tests and have not been independently verified, and the company notes they will vary by workload, network location and comparison method. Since Jev is available only through an early-access waitlist, the near-term question is whether early users find the constrained output reliable enough to embed in production, which could make this style of decision model a complement to, not a replacement for, the language models it is measured against.