TrailCurrent YouTube Code Samples

Code, firmware and CAD from TrailCurrent YouTube videos. Each folder is a separate project with its own README covering what it does, what hardware it needs, and how to build and run it.

These are demonstration projects. They are meant to be read, built and adapted, not dropped into a product unchanged.

Samples

Sample	Runs on	What it is
Desktop & Linux
TrailShell	Waveshare PocketTerm35 (Raspberry Pi 5)	A touch-first desktop shell for a 3.5" 640×480 screen with a built-in keyboard: status bar, swipe-down shade, launcher and task switcher.
LinuxTouch	Waveshare PocketTerm35 (Raspberry Pi 5)	A phone-style shell on top of full Raspberry Pi OS, with a home screen, one fullscreen app at a time and a card switcher. In development.
FNIRSI PSU Control	Linux, Windows, macOS	An unofficial cross-platform desktop app for FNIRSI DC bench power supplies (tested on a DPS-150), built by reverse-engineering the Windows-only vendor app. Beta.
Genie NPU installer	Radxa Dragon Q6A	A one-file script that runs Llama 3.2 1B on the Qualcomm NPU and serves it as an Ollama-compatible API. See below.
ESP32 displays
Off-Grid Power Dashboard	Waveshare ESP32-P4 7" touch	A battery, solar and tank dashboard showing the new features in EEZ Studio 0.29 and LVGL 9. It also runs in EEZ Studio's simulator.
Headwaters Map Viewer	Elecrow CrowPanel Advance 10.1" ESP32-P4	Walks through Wi-Fi setup and token sign-in, then shows map tiles from a TrailCurrent Headwaters gateway.
Rotary macro pads
Rotary Macro Pad, 2.1"	Makerfabs MaTouch ESP32-S3 Rotary 2.1"	A USB keyboard with a round touchscreen in a rotary ring. Its keys change to suit the app you're using: FreeCAD, Blender, Kdenlive, VS Code, GIMP, Inkscape, LibreOffice and browsers.
Rotary Macro Pad, 1.28"	Elecrow CrowPanel 1.28" Rotary (ESP32-S3)	The same macro pad, redrawn for a 240×240 screen.
Rotary Macro Pad, 1.46"	Elecrow CrowPanel 1.46" Rotary (ESP32-S3)	The same macro pad, redrawn for a 360×360 screen.
Crack the Safe	Waveshare ESP32-S3-Knob-Touch-LCD-1.8	A safe-cracking game you play by feel: the haptic motor clicks when the knob lands on each number, and the vault door swings open on glowing gold. In development.
ESP32 boards
ESP32-P4-WIFI6 Demo	Waveshare ESP32-P4-WIFI6	Wi-Fi setup from a phone, a web dashboard, microSD file management, microphone and speaker, a camera, and push-to-talk to the Peregrine voice assistant.
Antenna A/B Test	Waveshare ESP32-S3-RS485-CAN	Measures the onboard ceramic antenna against an external SMA antenna and reports the difference in dB.
CAD
Torsion Table CAD	FreeCAD + LowRider 4 CNC	CAM jobs for cutting the plywood ribs of a torsion box table.

The three macro pads share one companion program, which runs on your computer and tells the pad which app has focus. The pads identify as the same USB device, so one install works with any of them.

Getting the code

git clone https://github.com/trailcurrentoss/YouTubeCodeSamples.git
cd YouTubeCodeSamples

FreeCAD files (*.FCStd) are stored with Git LFS. If you want the CAD, install Git LFS first, or run git lfs pull after cloning.

Most samples need a specific toolchain, for example a particular ESP-IDF version. The sample's README says which. Start there rather than assuming the same setup works for every folder.

Genie NPU installer

install-genie-npu.sh sets up a local AI chat server on a Radxa Dragon Q6A. It runs on the board's Qualcomm Hexagon NPU instead of the CPU, at roughly 12 tokens per second.

Start from a stock Radxa OS R2 (or newer) image. Log in as the radxa user (not root) and run:

bash install-genie-npu.sh

It needs internet access and will ask for your sudo password. The script:

1. Installs btop (to watch CPU load) and jq (to read the streamed replies).
2. Downloads the NPU build of Llama 3.2 1B, about 1.3 GB.
3. Changes two settings so the server sits near 0% CPU when idle instead of keeping three cores busy.
4. Installs genie_server.py from TrailCurrent Peregrine as a systemd service that starts on every boot.
5. Waits up to two minutes for the model to warm up.

When it finishes, the server answers on http://localhost:11434 using the same API as Ollama, so tools that talk to Ollama can use it. The script prints a curl command you can paste to try it.

To use a longer context window, set CTX_LENGTH first, for example CTX_LENGTH=2048 bash install-genie-npu.sh. Only context lengths that Radxa publishes a model bundle for will work.

License

MIT. See LICENSE.