That is the part that makes it usable in a pipeline: you pick where the line sits, so a screening job can run strict and a user-facing warning can run loose, off the same call. Responses come back in milliseconds, which means it is cheap enough to run over every inbound message rather than the suspicious ones you already noticed.
A bilingual fraud-detection engine for text, reachable both as a REST API and as an MCP tool. It detects the language, routes to the matching model, and returns a risk judgement rather than a paragraph of hedging.
- A risk score, a verdict of fraudulent or normal, and a risk level of low, medium or high
- A fraud probability alongside the verdict, so you can set your own threshold and control false positives
- Bilingual handling: Chinese and English text are detected automatically and sent to their own models
- Detection combining a LightGBM model with Chinese keyword rules, aimed at scams, phishing, fake recruitment and gambling advertising
Registered as a remote server over HTTP at `https://fraud-detector-production-7cf8.up.railway.app/predict`, version 1.0.0, so there is nothing to install for the hosted path. To run it yourself, clone the repository. No credentials are documented for either route.