Laya Open-Source Decision Model Tested for E-Commerce Fraud Detection
A developer spent a weekend testing Laya, an open-source alternative to JEV, for detecting fraudulent orders in e-commerce — finding it faster and cheaper than large language models for simple decision tasks. Laya is built on ModernBERT and operates as a non-autoregressive System One model, returning scores or choices rather than generating text. Testing revealed that business approval logic should remain in application code, with Laya only estimating fraud risk, since letting the model make final decisions led to over-cautious results. The developer also found that phrasing input data in natural language reduced semantic bias, and that GPU hardware cut response times from over five seconds to under one second. The project is publicly available on GitHub, and the author notes a broader industry trend toward specialized, task-specific AI models over general-purpose LLMs.
