Jev is Type Safe AI's new model for turning natural language and application state into typed decisions. You supply context and define possible answers; Jev returns choices, scores, and probabilities as JSON. Part of a wider effort to make AI outputs code-friendly, Jev uses a new architecture, a parallel sampler, and Reinforcement Learning for Calibrated Decisions, avoiding token-by-token generation for speed and cost gains.

Type Safe's docs highlight patterns like speculative fan-out, confidence-gated routing, composite scoring, and intent routing, enabling probabilistic workflows. The post explores integrating Jev with Apache Arrow, which speeds data pipelines by avoiding JSON conversions. An Arrow schema represents Jev's answers: Choice, Noul, and Score types, each with specific fields like probabilities and confidence.

Arrow extension types attach semantic meaning, using structs and fixed-size lists for efficiency. Choice stores a byte index into shared labels; Score includes a legend. This design leverages Arrow's performance for structured data, potentially enhancing Jev's decision pipelines.