Dead Air: a fictional shortwave app that runs a small AI model on your own CPU

Dead Air is an open-source fictional shortwave radio that generates short transmissions from a place, an unsettling detail, and a mood the user supplies. The app runs SmolLM2-1.7B-Instruct, an Apache-2.0 open-weight model, locally in the browser-facing setup via Transformers.js and ONNX Runtime on the CPU, with no cloud fallback. Its server listens only on loopback, validates inputs, and limits each response to 140 tokens with one generation at a time. The author found the 1.7B model followed prompts more closely than the 360M version, though generation on their machine took about 28 seconds including model loading. The project is MIT-licensed and documents its limitations, including the model's tendency to repeat or cut off mid-sentence.