Built a fully offline suitcase robot around a Jetson Orin NX SUPER 16GB. Gemma 4 E4B, ~200ms cached TTFT, 30+ sensors, no WiFi/BT/cellular. He has opinions.

Sparky runs entirely on the Jetson. Gemma 4 E4B at Q4\_K\_M via llama.cpp with q8\_0 KV cache and flash attention. 12K context, native system role, sampler defaults from the model card. Cached TTFT around 200ms, sustained 14-15 tok/s. SenseVoiceSmall for STT, Piper for TTS with 43Hz mouth sync, Pixi