REFEREE

REFEREE came from a simple frustration: delivering high-resolution video at scale is expensive, and the compute needed to produce it already exists in hundreds of millions of consumer GPUs sitting idle while people watch TV. The idea of shifting that work to the edge — literally the user's own desk — and letting RTX Tensor Cores and Radeon AI accelerators do what they were built for felt like an obvious inversion of the standard streaming model. Choosing Rust for the backend was a deliberate decision for memory safety and execution speed; the pipeline needed to feel invisible, and that meant no bloat, no leaks, and reliable teardown after every session. The hardest engineering problem was not the AI processing itself — NVEncC, VCEEncC, and their respective AI SDKs handle that brilliantly — but building a vendor detection and fallback layer robust enough to transparently swap between NVIDIA SDKs, AMD SDKs, and open-source alternatives like ArtCNN ESRgan, RIFE, and Libplacebo depending on the hardware and OS available. Tauri kept the desktop app footprint minimal while Docker and headless server builds opened the door for anyone who prefers to self-host.