convaiinnovations/laya-demo

Laya Demo

Laya is a fast System 1 decision engine: send a state and typed questions, get typed answers with probabilities and a confidence score. It never generates text, so there is nothing to parse and nothing to hallucinate.

The tabs are the patterns people use most: support triage, email and phishing, LLM guardrails, RAG passage filtering, moderation, model routing, and a free-form playground. Every tab asks all of its questions in one pass and then decides what to do with plain code: the thresholds live in the app, not in the model.

Multilingual routing

The Multilingual routing tab runs laya.Router: the script of the incoming text decides which checkpoint answers, before the forward pass.

Routing is decided up front because the model's own confidence gives no warning when it cannot read the input. Across 51 languages on MASSIVE intent (20 options, random = 0.050) the English checkpoint macro-averages 0.227 and clears 3x random on 23 of 51 languages, while reporting 0.89-0.99 confidence throughout — on Khmer it scores 0.000 accuracy at 0.952 confidence. The multilingual checkpoint reaches 0.366 and clears 3x random on 45 of 51.

from laya import Router
router = Router()
router.predict({"body": "I was charged twice"}, questions)          # -> laya
router.predict({"body": "मुझसे दो बार शुल्क लिया गया"}, questions)   # -> laya-multilingual

This is a preview checkpoint (421M parameters, trained with pure reinforcement learning against proper scoring rules on public datasets). It is strong on routing, classification, moderation and guardrails, weaker on rubric scores, and it saw no email data, so treat the email tab as generalisation rather than a trained skill.

Built by Convai Innovations.