Laya a decision making system one model.

A developer built an order risk detector on Laya, an open-source non-autoregressive decision model based on ModernBERT, to test whether a small specialized model can replace large generative LLMs for e-commerce fraud detection. The developer found that asking the model directly for an APPROVE/REVIEW/HOLD business decision caused it to hold all orders, so the pipeline was changed to have Laya output a fraud risk score that the application maps to actions (under 30 approve, 30-70 review, over 70 hold). Raw field names such as "failed_payment_attempts: 0" triggered semantic bias toward risk, while natural-language phrasing improved results; latency was about 5 seconds on a laptop, under a second on a T4 GPU, and roughly 4 seconds per request on a 2 vCore, 4 GB VPS.

JEV is the new hot topic in the tech community from last week. So I decided to spend the weekend testing Laya open-source alternative to JEV for E-Commerce Fraud Detection. Here’s what I learned... Laya can be run locally on your device Most AI projects nowadays use generative LLM models like GPT or Claude for almost every task. Which can be a huge overhead and costly for simpler and quick tasks. For simple decision-making tasks like fraud detection, support systems, validation we don't always need a huge LLM. We need something fast and consistent for decisions. That’s where Jev/Laya comes in. Jev/Laya is a decision model built on ModernBERT. JEV/Laya is a non-autoregressive System One decision model. Instead of generating text like a chatbot, it takes structured information and gives answers as scores, choices or nouls. To understand and test it, I built a small Order Risk Detector. Here are the takeaways after doing this small project At first, I asked Laya: "Should we APPROVE/REVIEW/HOLD this order?" It wanted to HOLD all orders even when the orders looked completely normal. The problem was that "Approve/Review/Hold" is a business decision. Every e-commerce company can have its own rules for this. Like a company selling very expensive items can be more cautious. Whereas a small company can ignore some indicators and approve orders easily. So I changed the approach: The model estimates the fraud risk and our application decides what to do. For example: if Risk < 30 → Approve Risk 30–70 → Review Risk 70 → Hold I also found something interesting. When we gave the model raw data like failed payment attempts: 0 the model focused on the word "failed" and treated it as a risk signal. It's a classic semantic bias issue also on LLM's. So I changed it to a more natural one "Customer has a clean payment history with 0 failed attempts." or you can also use "all payment attempts: successful" This gave much better results. We also tested how fast Laya runs on different hardware. On my potato laptop, one request took around 5+ seconds. On a T4 GPU Used from Colab , the response dropped to less than a second. This is the actual power of these kinds of models. After testing, I moved the Laya-serve backend to my own VPS server, so the model is now running on my own server with 2 vCore and 4 GB Ram. Each request takes 4s on average. The experts are also predicting that the industry is shifting towards specialized models and task specific tools. Check out the project here: Github: https://github.com/Shariful-Islam-Sourav/order-risk-management-laya https://github.com/Shariful-Islam-Sourav/order-risk-management-laya Site: https://order-risk.netlify.app/ https://order-risk.netlify.app/