I built a chakra detector for the new ESP-Mosaico (ESP32-S31): strike a singing bowl and the whole screen, plus your Hue lamps, takes its colour
Moderator Announcement
Read More »
Your submission has been temporarily paused because you did not select a post flair.
To keep our community feed organized and ensure your post reaches the right experts, you must tag your post. (If you are on mobile and do not see the categories, scroll down inside the flair window!).
How to fix this right now:
Click the three dots (…) or the mod shield on your post.
Select Change Flair or Add Flair.
Choose the category that matches your topic (e.g., Software Help, Hardware Help, I Made a Thing!).
Once you assign a flair, your post will automatically re-enter the moderator queue for approval. Remember that project showcases are limited to one post per project per 90 days.
I am a bot, and this action was performed automatically. Please contact the moderators of this subreddit if you have any questions or concerns.
Preatty cool project. And also super fascinated by mosaico: how did you use ai on it?
Thanks! Mosaico is built for this kind of workflow. The board’s base firmware (Vibe Mode) speaks a protocol called ESP-Iris, so an AI agent can talk to the device: install an app, update it, read its state.
I used Claude Code for pretty much all of it:
Detection: it wrote the pitch detection in C, then tuned it against real bowl recordings.
UI: it designed the screen and checked every change in the GSP simulator with my Mac’s microphone.
Testing: it measured the latency from strike to colour, which is where Flow and Strike came from.
Then I wanted to know whether AI actually helps, so I had it run 33 freely licensed recordings (bowls, gongs, drums, rainsticks, shakers) with and without an AI sound gate:
Without AI: bowls light up 90% of the time they sound, as they should. Drums, rainsticks and shakers never trigger a colour, because the “steady tone” rule already filters them out. The real problem is gongs: they light up 88% of the time.
With AI: a small EfficientAT model, plus a tiny head trained on these clips (tested on clips it never saw), brings gongs down to 18% while bowls keep 86%. With a 2 s window gongs drop to about 3%, at the cost of reacting a second later.
So the app now has an “AI sound gate” switch. Only bowl-like sounds change the colour, transitions get softer, and a gong or drum hit gives the light a short brightness kick instead of a wrong colour.
The model is about 1.9 MB as int8/int16 and should run on the board’s AI instructions in roughly 30 to 60 ms per second of audio. Next I add Video AI detection to see if it detects the right Chakra Bowl from the Video…
We removed this post as it is just an ad without providing any additional benefit to this community