Laya
title: laya type: tool created: 2026-09-20 last_updated: 2026-09-20 related: ["jev-ultrafast"] sources: ["https://github.com/NandhaKishorM/laya"] radar_quadrant: Tools radar_ring: Assess radar_position: center
Agent pipelines often spend slow, costly text generation on decisions that are really classification: which department should handle a ticket, how urgent it is, whether a user threatens to leave. laya is a decision engine that answers such typed questions in a single forward pass without generating any text, so there is nothing to parse and nothing to hallucinate.
A caller supplies a state (text, an email, a ticket or a JSON document) and a set of typed questions. The three question types are choice (a label with probabilities), score (a level on an ordinal rubric) and noul (a calibrated probability that a statement is true). The README reports 33 ms for one question and 7.2 ms per question when batched, measured on a T4 GPU. It also ships presets for model routing, prompt guardrails, content moderation and support-ticket triage.
Three checkpoints are provided: an English model, a multilingual model covering 100+ languages and a typed-decisions model. A built-in Router detects the script and language and dispatches to the right checkpoint for each request.
The README documents that the English checkpoint collapses on non-Latin scripts, for example 0.000 accuracy on Khmer at 0.952 confidence, so confidence gating alone cannot catch the failure. It also reports that without preloading, alternating languages forces a model reload of 7 to 10 seconds per switch. All benchmark figures are the project's own.
The repository is Apache-2.0 licensed and was created on 2026-09-18. It had about 2,350 stars and 21 open issues when checked. The README compares itself against TypeSafe Jev, the decision service behind jev-ultrafast.
laya is placed in Assess at the center position. A generation-free decision step is an attractive idea for routing and triage, and adoption has been fast, but the repository is days old, the benchmarks are self-reported and there is no first-person use.
