This is my space for collecting some interesting projects I have worked on. Enjoy -
Will Byers LED Chat: Talk to the Upside Down
I originally built this project in May 2022 for Stranger Things Season 4, recreating the iconic scene where Will
communicates from the Upside Down through Christmas lights. With Season 5 approaching in 2025, I've revived the project
with Claude AI playing the role of Will Byers, responding to your messages by flashing individual letters on a wall of LEDs.
The hardware consists of a Raspberry Pi 4B controlling a strand of 100 individually addressable WS281x LEDs connected
to GPIO 18. Each LED is positioned above a letter of the alphabet painted on a piece of cardboard, cutout, and taped to the wall.
The system runs on a 12V power supply for the leds (5v is available) and a separate USB power supply for the PI.
The major bill of materials includes:
Raspberry Pi 4B (any Pi with GPIO works)
12V WS281x LED strip (100 LEDs)
12V power supply for LEDs
Logic level converter (3.3V to 5V)
To create the display, my daughter hand-painted all 26 letters of the alphabet on cardboard, we cut them all out, then hung them on the
wall. Since we can control any light, I held off on creating a letter-to-light address map until after everything was hung.
In this way I could illuminate the lights individually with the code, and then note which address was closest to which letter.
Now with AI technology exploding since 2022 - (especially language modeling and comprehension), I resurected the project just in time for
Stranger Things Season 5!! The updated project is available at github.com/kyjohnso/will_byers
and includes two demos: a static message player that displays pre-programmed text, and an interactive chat mode where
Claude AI responds in character as Will Byers. The AI integration uses the Anthropic API to generate contextual responses,
with each letter flashing in sequence on the wall—complete with dramatic pauses for spaces, just like in the show.
Coming Soon: Before Season 5 premieres on Thanksgiving 2025, I'll be adding voice activation using
modern speech-to-text technology—finally realizing my original 2022 vision of Alexa integration (similar to my
Alexa-enabled Lego Mindstorms maze solver below), but with full conversational AI instead
of simple voice commands.
I would LOVE to see a bunch of posts of you all recreating this scene in the run up to the finalle - this is going to be a good one!
Regolith Voxel
Regolith Voxel is
an experimental mining simulation game that serves as a testbed for reinforcement learning research. The project explores
how RL agents can learn to manage complex resource extraction operations through indirect control of semi-autonomous mining equipment.
The core concept investigates indirect control paradigms where players provide high-level strategic direction
rather than direct commands. Each piece of mining equipment operates autonomously, making local decisions based on its type,
position, and observations of the mineral-rich environment. This creates emergent behavior as multiple agents coordinate
their activities—samplers scout for resources, miners extract materials at different depths, refiners process raw ore,
and transport units move goods between sites.
What makes this particularly interesting for RL research is the network effects between agents. A single
agent's optimal strategy depends heavily on what other agents are doing: transport units need miners to produce ore,
refiners need raw materials delivered, and efficient sampling can guide mining operations to high-value deposits. This
multi-agent coordination challenge, combined with the hierarchical organization of equipment and the spatial dynamics
of resource management, provides a rich environment for exploring how reinforcement learning agents can learn to cooperate
and adapt in complex systems.
Built with Rust and the Bevy game engine, the project features procedurally generated mineral maps, a drag-and-drop
equipment hierarchy system, and a modular architecture designed to integrate RL training loops. Future work will focus
on implementing policy networks for autonomous equipment behavior and experimenting with reward structures that encourage
efficient multi-agent coordination.
SkySplat: 3D Gaussian Splatting Toolkit
SkySplat is a comprehensive Blender addon
that streamlines the creation of photorealistic 3D Gaussian Splats from drone footage and aerial photography. The project
bridges the gap between raw video capture and finished 3D scenes by integrating structure-from-motion (COLMAP),
Gaussian Splatting training, and Blender's powerful 3D workflow into a single cohesive toolkit.
What sets SkySplat apart is its focus on coordinate system transformation and spatial integration. Rather
than treating reconstructed scenes as isolated assets, the addon allows you to rotate, scale, and position COLMAP point
clouds within Blender's native coordinate space before Gaussian Splatting training. This means you can align reconstructions
with reference imagery, combine multiple splats, or integrate traditional 3D models into photorealistic captures—all while
maintaining accurate spatial relationships.
The complete workflow runs entirely within Blender: import video, extract frames with metadata, run COLMAP reconstruction,
transform and orient the scene, train Gaussian Splats with the bundled Brush renderer, and load the results directly into
your viewport. The addon bundles pre-compiled Gaussian Splatting binaries for all platforms and provides real-time training
visualization, making advanced 3D capture accessible to artists and researchers without complex dependency management.
I presented this project at BlenderCon 2025, demonstrating workflows for architectural visualization, terrain reconstruction,
and mixed CGI/photogrammetry compositions. The project is open source on GitHub at
kyjohnso/skysplat_blender,
with comprehensive documentation and example workflows at skysplat.org.
ThreeJS Map
This is the current state of a very simple map and terrain visualization in threejs. It will accept latitude and longitude
and zoom level and render a single tile in the canvas. The terrain data is provided by
Mapzen documented at
tilezen under the
Creative Commons BY 4.0 license.
The topo map is provided by
Open Topo Map under the
Creative Commons BY-SA 4.0 license.
My code can be found at
ThreeJS Map.
Try playing around with the "random location" button.
In February-2025, I completed Blue Dot Impact's AI Alignment course.
In this course we are exposed to a wide variety of AI alignment techniques, ideas, and research approaches. In the final third of the course, I completed
a project in Computable Policy and World Model integration. You can find the post for this project at
Hylaean Sea. As with all of these projects (it seems) I plan to develop this
concept more in the future.
I have also been engaging a local chat group on some high level questions about AI Alignment Philosophy and will be including those prompts here as well.
Distribution of AI
A thought I have been pondering recently: assuming the economic, leisure, and cultural benefits of AI are very large (they may not be but lets assume they are), what are the major tipping points or key factors that will determine if these AI resources are concentrated in the hands of a few AI elite, or the AI upper class; or are spread out in society for everyone's benefit? Is it education, fear, use in businesses, open source models, or something else that could make the difference between it being available to the masses vs locked away?
AI Unlearning
There is significant research into AI alignment at the fine-tuning stage. Several techniques attempt to make LLMs forget, or unlearn harmful
knowledge that the base model acquired in the training phase on bulk internet data. We try to make it "unlearn" how to make a bomb, or how to
make a nerve-agent for example. Is this wise? Don't we want LLMs to have harmful knowledge so that they can recognize it, as it comes up in
prompts? Does this have parallels for human knowledge? Is there "harmful knowledge" that we should try to unlearn as a society?
Bootstrap Training
If we need humans to generate training or fine tuning data to train advanced AI's can they ever become more capable than humans?
Can a human train something more cognitively capable than itself? If so, how do we make sure that these more capable AIs are aligned with our goals?
Blender Con 2024
I was fortunate to have my talk selected for Blender Con 2024. I expected to have just a 20 minute talk but I was slotted for a live 50-minute walkthrough of some of the projects that appear on this page.
I began with a walkthrough of the geometry nodes approach to modeling parametric surfaces, then demonstrated how to use modifiers and scripting to model the 3D dodecahedron moon.
All 12 moon .stl files are available under a Creative Commons license at Printables.com.
Ideal Gas Simulation
My daughter was studying the states of matter in her science class and she had a fairly good grasp of the molecular (non-quantum) mental model of matter and how it translates
to the physical properties of the material in different states. She, however had a bit of trouble with visualizing why an ideal gas will take the shape of its container. After much
explaining, pretending we were molecules and dancing around, and throwing fruit off the walls and ceiling of the kitchen, we decided to have chatgpt help us visualize this in
a web simulation. The simulation was written in threejs and can be found at Ideal Gas Simulation
Dodecahedron Moon
This is a 3D printed design of the moon (real topography from
NASA Moon Kit) broken
into the curved hexagonal faces of a dodecahedron. The whole design is held together by
force-fit magnets. The full 12 STL files will be posted soon at Printables.
Frequency Difference
An exploration of the intersection of quadrics. Frequency difference (doppler difference) from a location to two different moving sensors
can be modeled as two intersecting cones with related cone angles. Set to the music of Hans Zimmer, it can be quite mesmerizing.
CGI Sandbox
An exploration of CGI with drone footage. So far done only with Blender motion tracking. Hopefully evolving over the coming months. Drone footage courtesy of my buddy Greg.
Intersecting Hyperboloids
A visualization of two intersecting hyperboloids with one of their foci shared. It is interesting to note that the intersection pencil will always intersect the plane formed by the three foci at a right angle.
Time Difference Visualization
A visualization of a time difference surface made in Blender.
Alexa-Enabled Maze Solving Robot
Alexa Enabled Maze Solving Robot - An entry into a contest to link Amazon Alexa Gadgets to Lego Mindstorms. I made a robot that you gave instructions to via Alexa to help it solve a Lego maze.
