TypeSafe has launched Jev, an artificial intelligence tool designed to produce structured decisions rather than conversational text. The company says the system works with a fundamentally different architecture from large language models and is aimed at decisioning tasks rather than chat or code generation.
Founder Diogo Almeida said the project began from the view that existing models were missing elements that would limit progress toward human‑level performance. TypeSafe claims Jev is almost 200 times faster and 450 times cheaper than competitors such as ChatGPT and Claude when applied to routing and classification tasks.
Jev accepts a state and typed questions, and returns choices, scores or yes‑or‑no probabilities with an attached confidence value. TypeSafe says this constrained output both reduces compute and prevents hallucination because the model cannot produce unconstrained prose.
The company positions Jev for high‑volume pipelines, real‑time applications and tasks where users need a single decision rather than explanatory text. TypeSafe acknowledges that traditional LLMs remain useful where open text, coding agents or easily checked correctness are required. Jev is available in early access and the vendor has removed its waitlist to expand developer access.
Security, Adoption And Industry Reactions
Adoption moved quickly as integrations appeared across vendor tooling. Vercel reported rapid uptake among teams. Cloudflare, LangChain and Langfuse added integrations soon after launch, and TypeSafe began offering new users a small free credit to experiment with the service.
TypeSafe published performance figures showing end to end latency between 70 and 500 milliseconds. Pricing cited in partner documentation lists input token costs at $0.042 per million, with free output. Those economics make automated review inexpensive compared with human checkpoints, TypeSafe and partners noted.
Security vendors and tool authors raised immediate warnings. Pydantic and TypeSafe both say Jev treats the state as data not as hostile. Pydantic warns that the order of options or enum members can change answers and that adversarially crafted content can steer responses.
An Octomind engineer published a test where an injected fake tool output changed Jev’s block probability on a dangerous command. The demonstration showed block probability and confidence falling when a planted field suggested preapproval. TypeSafe and Pydantic cite similar sensitivity in their limitations documentation.
LangChain shipped middleware that narrows what Jev sees and excludes tool output so fetched content cannot authorize its own execution. The middleware also recommends pairing automated decisions with human approval where consequences are high, combining deterministic checks and human oversight.
Security experts warn that decision models like Jev complicate existing agent identity and audit frameworks. Kayne McGladrey has tracked gaps in how enterprises assign identities and permissions to agents. Those gaps can widen permission sprawl when automated decision services inherit shared identities.
Enterprises experimenting with agent pipelines report keeping humans in the loop for high‑impact actions. Ivanti built short‑lived agent harnesses and required human sign‑off on agent outputs, and said reviewers caught invented details that agents proposed. TypeSafe and partners stress that Jev’s confidence is a margin rather than a literal probability of correctness.
