TypeSafe AI's Jev is a decision model designed for agentic security workflows, answering fixed-choice questions with probabilities instead of generated text. In a proof of concept for alert triage, Jev filtered 15-33% of alerts with 98% accuracy while reducing latency and cost by handling multiple decisions in parallel at 70x cheaper per token than frontier LLMs.
Alex built ChatJEVs, a text generation system using 521 Jev decision models that produces language through probabilistic choices rather than traditional neural network training. The system samples words by querying a frozen decision engine across multiple architectural versions, with each word generation requiring multiple batched requests to answer questions about grammar, context, and appropriateness.
Kev is a family of small decision models (0.8B–9B parameters) built on Qwen3.5 for classification tasks like customer support ticket routing. Users can run pretrained weights locally on CUDA or Apple Silicon, with a TypeSafe-compatible API and web playground for testing.
RoboKrunch benchmarked TypeSafe's Jev decision model on simulated warehouse robot fleet triage, finding it costs $24.57 per million decisions with sub-second latency and requires no training data. Self-hosting a small model becomes cheaper above ~1M decisions per month, but Jev wins on startup cost and simplicity for smaller deployments.
Jev, a TypeSafe AI decision model, evaluates structured state against typed questions and returns choices, scores, and probabilities without text generation. The AI SDK's experimental_evaluate API exposes Jev through the Vercel AI Gateway in TypeScript, enabling applications to route decisions—like support ticket departments or severity scores—based on model assessments while keeping business logic in code.