Over Engineered Light Bulb

back

If it ain't broke

don't
fix it.

Isn't this image so satisfying?
(o´▽`o)

View GitHub repo here

I have a huge collection of music and wanted to make the listening experience in Navidrome better and more immersive. So I used a Philips WiZ light bulb and designed a system that smoothly transitions my room lighting between the 5 most dominant colors of the currently playing album art to keep the experience intimate and cohesive.
Somehow, that led me into the embedding model rabbithole. Soon enough, I had a system where I could type a prompt like "soft relaxing bright" or "campfire" or "deep sea" or anything at all and my room's lights would change to that. I learnt a lot of things like, why bigger model is not always better, etc.

Table of contents →

Navidrome
The lighting
Semantic Search Setup
Semantic Search Implementation
YouTube Demo

Navidrome

I was checking my PS Vita's app store and came across this music streaming client, Nekodrome by NekoMimi which looked cooler than the Jellyfin client on PSV. That's when I thought I needed a Navidrome server. It took me some time to understand how it works and how I can make it truly mine.
The web client was working, the PSV client was working, Android client was working. I soon discovered a problem where the metadata and artist tags were not perfect. And some albums missed Album Art. It reminded me of MusicBrainz Picard as I had seen it in GSOC organizations list. With Picard and Navidrome, one can essentially host their own music streaming service with no ads or low quality music.
I carefully added Metadata to all my music using Picard and got Supersonic on my PC. I saw the blur in the background and thought, "How cool would it be if that background leaked outside the monitor...".

MCR's 'I Brought You My Bullets, You Brought Me Your Love' album playing on Supersonic with a background blur

It's good that I'm immune to seizure-inducing graphics or else i'd be
(°o°:;)

It was time to put my WiZ Light Bulb to some use and finally making a project other than, "Oh look, my room only has light when I am typing on my keyboard. This motivates me to type more and more". I setup pywizlight in a fresh virtual environment. I also checked how Navidrome's API works and played around with cURL a bit. I was able to pipe it through jq and get track name, etc and write a shell script that could do something like this below. Check out the code here.

fetch script like neofetch that shows currently playing album along with album art

The Lighting

I used the dotenv library to setup the variables like so. There are 2 modes in the script. Mode 1 ends the lighting show after the current album stops playing . Whereas Mode 2 dynamically checks everytime album has changed and crossfades the colors.

dotenv variables, like navidrome passoword, light bulb ip address, brightness and finally a 'mode' fror lighting

Hues are extracted using pillow. First, the Album-Art image is reduced to 64x64. Then the colorspace is changed to the HSV format. I intentionally reduced the hue resolution and cranked up saturation and brightness, so that I can get sharp colors instead of having to deal with 100 different shades of red. Now, I can have upto 32 colors as input.

32 colors here does not limit me to 32 light colors, I have a trick to get the full 16.7 million colors (^_-)

code snippet showing how the color extraction works, it puts all the colors into 32 buckets and returns the top 5 most populous buckets

Once we get the top 5 colors, we can use a simple for loop and go from lowest to highest (like a gradient -> to get all colors instead of just 32) of the 5 selected hues by adding small increments. While converting hues back to RGB (that's the format pywizlight works with) I hardcoded the Saturation and Brightness values to 1.0 because I did not want the album art to control brightness. I have used BRIGHTNESS in the .env file so that I can manually set the brightness or control it based on time of day or other external factors. Letting the album art decide the Brightness value is kind of like a missed opportunity in my opinion.

code snippet containing the loop for gradient light

It takes 60 seconds to go from hue n to hue n + 1. pyqizlight uses the asyncio library. Thanks to it, I was able to learn about non blocking loops, or else it would have been like my Voice Assistant project where everything was being blocked because it was single threaded.

And that's how I was able to get reactive lights based on currently playing album.

Semantic Search Setup

I realised I had one embedding model in Ollama I never used: nomic-embed-text. I looked around duckduckgo for a while and came across color-names. It is a list of colors with the name of the color being something a human would say, like the ones below. I knew this was perfect for semantic search.

So... uhh....
A thesaurus that needs a graphics card
(ㆆ_ㆆ)

I tested semantic search with just the key value pairs from colors.json, but the results were underwhelming. So I decided to add another layer of computa- ahem, intelligence to the program. My plan was something like this:
We get a tiny LLM to take the color names as input and generate 5 other words/phrases that could match it closely. Then we find the embeddings of these. This would make it drastically easy for the embedding model to match an input prompt to any of the colors.

Semantic Search Implementation

I searched ollama's model directory for newer and smaller models. I found a few that would fit my use case. all-minilm, and snowflake-arctic-embed and IBM's embedding model: Granite. I went on to benchmark some of the Embedding Models (I call it the great Embedding Model BakeOff May Week-4 2026 Edition to see which of these cooks my PC first).

Model	Time Taken	Quality (opinionated)	Notes	Results
granite:33m	00:16:03	linkedin 🧠📉	i would use rapidfuzz or pywal instead, like a caveman Truth be told, it seems like this is better for RAG than creative tasks. Still very 🧠🪦	check here
all-minilm:22m	00:09:52	excellent	it's better than 33m variant, but feels robotic	check here
all-minilm:33m	00:11:13	good	no soul + a bit corpo	check here
snowflake-arctic-embed:22m	00:08:07	excellent	perfect model for this use case	check here
snowflake-arctic-embed:33m	00:09:36	good	loses character compared to smaller model	check here
larger models nomic-embed-moe, bge-m3, mxbai-embed-large	1-2 hours	bad	not really good for this use case	NA not gonna waste any more time

I just like Qwen alirght?
Definitely not sponsored by Alibaba...
shhh...

I also benchmarked a few qwen LLMs as well. Below are the results of the LLM BakeOff May Week-4 2026 Edition:

Model	Speed	Quality	Notes
qwen3.5:0.8b	medium	Best overall	Good creativity
qwen3:4b	slowest	Straight up bad	"aktually 🤓☝️" and harder to follow rules
qwen3:1.7b	fast	Cinematic	Very good but always "soft glow", "ethereal", "muted tones"
qwen3:0.6b	fast	Repetitive	Same as 1.7b but obsessed with "ethereal"
qwen2.5:0.5b	fastest	Alright	Very basic model, just sticks to catchphrases like "ethereal"

These models actually got close to baking-off my PC, even 1 generation was taking close to 20 seconds sometimes, and I had to do this for 31k colors. I did the "smart thing" and ran it for a few demo prompts only. Therefore, the results may not be perfect. Upon that, I believe my prompt engineering could be better. I got ChatGPT to make this prompt for me because I'm not that good at talking to small LLMS. Generating tags for each of the colors using Qwen 3.5: 0.5b took approximately 6 hours.

Prompt to run it locally on my 3GB vram GTX 1060.

Now that the LLM had generated ~5 (excluding negative prompts which were filtered) extra tags for each of the 31k colors, I ran a loop, found embeddings for all of them and saved them in numpy verctor arrays, because we gonna do some vector math . This process took approximately 8 minutes.

Code snippet showing embeddings getting appeneded into the numpy array

Now, I just have to make a script where I ask for a mood and the python script does the semantic search in the numpy vector array and ranks the top 8 colors with the most embedding matching. It first generates an embedding for input text, then searches the array using cosine similarity.

So it finds the "direction" of the input phrase.
Get it? Get it?
(✖╭╮✖)
Ugh... nevermind

After this, I simply use the same Gradient Light from earlier and TA-DA!!

Youtube Demo

A few screenshots of the colors being generated by the semantic engine.

example of colors generated for: cyberpunk bright neon

example of colors generated for: cyberpunk bubblegum

example of colors generated for: blood pain anger

example of colors generated for: deep sea

example of colors generated for: green field with blue sky

Yet to record, edit and upload video

Github Email ORCiD