Anil-matcha/awesome-jev-by-typesafe

What it solves

Jev is designed for software that requires structured, typed decisions rather than generative text. It bridges the gap between brittle, rule-based systems and expensive, unpredictable generative LLMs by providing constrained answers and probability distributions that software can use for branching logic, routing, and gating.

How it works

Jev operates as a "System One" model that evaluates a provided state (text or JSON) against a set of typed questions. Instead of returning a paragraph, it returns specific primitives:

• Choice: Selects one value from a predefined, unordered finite set (e.g., intent classification).
• Score: Measures a position on an ordered rubric (e.g., severity or frustration levels).
• Noul: Estimates the probability (0 to 1) that a specific statement is true (e.g., detecting prompt injection).

Developers integrate Jev via Python or JavaScript SDKs, using the answers and associated confidence levels to trigger deterministic code paths, human review, or fallback mechanisms.

Who it’s for

Software engineers and AI architects building production-grade automation, RAG pipelines, and agentic systems who need reliable, structured outputs for control flow, safety guardrails, and high-speed classification.

Highlights

• Typed Primitives: Provides Choice, Score, and Score for deterministic software integration.
• Confidence-Gated Routing: Allows developers to set thresholds for automation versus human fallback based on probability distributions.
• Speculative Fan-out: Enables asking multiple independent judgments in a single API call to reduce latency.
• Broad Application: Useful for support triage, RAG passage filtering, semantic re-ranking, and LLM guardrails.
• High Efficiency: Reported response times of 70–500 ms, making it significantly faster than general-purpose generative models.