Jev: Fast, Typed AI Decisions with System One Models

Course Overview

This course introduces Jev, a System One model designed to replace slow, hallucinating large language model (LLM) calls with fast, typed decisions for tasks such as routing, triage, and guardrails. Jev returns typed values with calibrated probabilities in milliseconds, enabling cost-efficient and validated AI decision-making. The course covers the three primitives (Choice, Score, and Noul), confidence gating, production patterns, engineering for production, and evaluation of AI decisions, culminating in building an end-to-end support-desk autopilot system.

Key Takeaways

• Model in-app decisions as Choice, Score, or Noul questions over structured state.
• Gate real actions based on calibrated confidence rather than single labels.
• Apply five documented production patterns to optimize AI decision systems.
• Run Jev inside request handlers with retries, rate-limit handling, and latency budgeting.
• Validate thresholds using gold sets and calibration checks.
• Reduce routing and triage costs by approximately two orders of magnitude compared to frontier LLMs.
• Recognize and design around nine documented failure modes.
• Build a support-desk autopilot that reports cost per decision.

Prerequisites

• Comfortable writing Python (3.10+).
• Experience calling an LLM API.
• Access to a TypeSafe early-access API key for hands-on labs.
• OpenAI or Anthropic API key for cascade labs.

Target Learners

• Engineers running LLMs in production concerned with latency and cost.
• Professionals involved in triage, routing, classification, moderation, extraction, or re-ranking using chat models.
• Backend and platform engineers needing decisions within 200ms instead of several seconds.
• Technical leads deciding AI placement within request paths.

Final Project

Build Fernway’s end-to-end support-desk autopilot system that gates actions by justified thresholds, combining deterministic code for routine cases and LLMs for tail cases. The project includes evaluating and reporting per-ticket cost against a baseline.