Developer Creates In-Browser Chrome Extension to Filter YouTube Feeds Using Local WebGPU AI
Developer Manjunath Shiva released opendecider, an open-source Chrome extension that filters YouTube video feeds locally using an in-browser ONNX decision model accelerated by WebGPU. The extension runs entirely on-device without external server dependencies or API keys, hiding user-specified content categories and YouTube Shorts offline after an initial model download. This project highlights the growing viability of client-side AI inference, protecting user privacy while eliminating recurring backend GPU costs for developers. It demonstrates how lightweight ML models combined with WebGPU can bring responsive, personalized content filtering directly into everyday web browsing. Powered by ONNX Runtime Web, the fp16 WebGPU model (755 MiB) evaluates 40 video titles in 1.1 seconds, achieving a 0.934 balanced accuracy on custom filter rules. It includes a 569 MiB q8 WASM fallback for CPU processing, automatically unloads from memory after 10 idle minutes, and features a right-click tool to check text for prompt injection.
## BACKGROUND
WebGPU is a modern web graphics and compute standard that gives browser applications direct, high-performance access to hardware GPUs. ONNX Runtime Web allows developers to execute machine learning models inside JavaScript applications using hardware-accelerated WebAssembly or WebGPU execution providers.