Summary
Sam Witteveen builds a working model router using Jev, a “system one” decision model from TypeSafe AI that returns typed, probability-scored answers instead of generated text. The router classifies incoming requests and sends each one to the right model: simple chitchat goes to a locally hosted MiniCPM 5.2B model, coding tasks route to DeepSeek V4.1 Flash in the cloud, and image requests go to a locally run Qwen Image 2.1 model that had just been released.
The video explains Jev’s three question types — choice, score, and null — and shows how each maps onto real routing decisions: choice handles categorization, score estimates task difficulty to decide if a small model can handle it or needs escalation, and null checks for sensitive data like API keys so private information never leaves the local machine. Live demos show Jev making routing calls in roughly 300 milliseconds at a fraction of a cent per request.
Aimed at developers building cost-efficient, privacy-aware LLM pipelines, the video demonstrates a practical alternative to routing every request through an expensive frontier model, using cheap, fast decision-making to triage traffic across local and cloud-hosted models.
📺 Source: Sam Witteveen · Published September 21, 2026
🏷️ Format: Hands On Build
