I built a Christmas tree that answers questions using light. It runs locally on a Raspberry Pi 4 (8GB) using WS2812B addressable LEDs and a small interface.

I used a single prompt that makes the model return a structured JSON response, which made it easy to map yes/no/maybe and emotions directly to LED colors and patterns.

I tested multiple LLMs and eventually settled on Gemma3:1B running locally on the Raspberry Pi. It performed surprisingly well both in terms of reasoning quality and computational efficiency. I would really appreciate any suggestions for improving performance or adding creative features.

Happy holidays!