hardrave/NIGHTRUN ? reverse-engineered prompt

Build me this weird little project that boots straight into a local LLM from a USB stick, with no normal operating system underneath.

I want it to start from firmware, show a simple fullscreen chat interface, load a quantized model into RAM with checksum verification, and then let me type prompts and get responses completely offline. It should work on ordinary x86 64 UEFI PCs, and also on a Raspberry Pi 5 if possible. Make the installer guide me through choosing a target, downloading a supported model, building a bootable image, and flashing it safely to removable media.

Please keep it robust and cautious, especially around disk flashing and model loading. I’d like live status while the model loads and while it is generating text, plus basic keyboard input and scrollback. If there are model or tokenizer details from the official docs that matter, look them up online and match them closely.