ai lab
Also indexed as typesafe-ai · typesafeai · jev
TypeSafe AI
TypeSafe tests whether applications can delegate frequent, bounded decisions to a specialized model while reserving generative models for explanation and open-ended work. Jev makes confidence and structured outputs central to that interface; independent evaluation is still needed to assess its reliability and performance across deployments.1,2
Profile
Overview
A specialist decision-model company
TypeSafe AI is a specialist model company whose team page names cofounders Diogo Almeida, CEO; Sasha Sheng, COO; and Erik Gafni, CTO. Its public model program focuses on structured decisions for software workflows and agents. The reviewed sources establish the team and product, but do not establish a founding date or verified financing total.3
Jev and RLCD
On September 15, 2026, TypeSafe introduced Jev in early access as its first System One model. The company calls its training approach Reinforcement Learning for Calibrated Decisions, or RLCD. Jev accepts text or JSON state and answers structured questions with choices, scores, and confidence. It does not generate free-form prose. The documented API uses the mutable alias jev-latest; an immutable release identifier and context limit were not verified.1,2
Evidence and limits
The launch compares Jev with larger language models on company-designed workflows and advertises substantial latency and cost reductions. These are publisher claims, not independently reproduced benchmark results. Some workflow comparisons use other models' answers as a reference rather than established ground truth. Schema-constrained output also should not be mistaken for proof that decisions are always correct.1
Notable contributions
- 01Notable contribution: a structured-decision APIJev exposes choices and numeric scores with confidence through the System One API. This describes the released interface, not a claim that TypeSafe invented calibrated decision models.2