The author details the development of Laya, a multilingual System 1 decision engine achieving 32.8 ms latency on a single GPU, positioning it as a faster, open-source alternative to proprietary models like Jev from Type Safe AI. The author claims prior art dating to March 2025 (arXiv:2503.23303), including models on Hugging Face, datasets, and a PyPI package, predating Type Safe AI's September 2026 launch. Founded by Diogo Almeida, a ChatGPT co-inventor at OpenAI, Type Safe AI introduced RLCD (Reinforcement Learning for Calibrated Decisions) at $0.042 per million tokens with 150 ms latency.

Laya addresses the bottleneck of using generative LLMs for simple classification tasks (routing, spam detection, urgency scoring), which adds latency (500–2000 ms), cost, and hallucination risk. Instead, Laya uses bidirectional encoders for instant, calibrated probability outputs via three primitives: choice (categorical selection), score (ordinal rubric placement), and noul (null/abstain detection).

The model family supports over 100 languages, operates at 7.2 ms/question batched, requires zero API subscription, and releases weights under Apache 2.0. The author frames Laya as the culmination of fixing architectural limitations in prior work, moving from vertical sales trajectories to horizontal, generalized schema decisions.