← Back to Projects

VisionNode

VisionNode was a real-time facial recognition project built around a split embedded architecture: an ESP32-CAM handled image capture while an STM32 took care of control and display logic.

What I wanted from it

I wanted to build something that felt like an actual embedded systems project rather than a machine learning demo glued to a laptop. The point was to make lightweight inference live on constrained hardware and then move data between devices cleanly.

Stack

• ESP32-CAM for image capture.
• STM32 for controller and display responsibilities.
• Edge Impulse and TensorFlow Lite for training and deployment.
• I2C and UART for communication between modules.
• MbedOS for the embedded side of the system.

What I learned

• Keeping responsibilities split across devices makes debugging and iteration much saner.
• Embedded ML gets interesting when latency, memory, and communication constraints are all part of the problem.
• A project like this becomes much more about systems design once multiple peripherals and protocols are involved.

Why it matters to me

VisionNode is still one of the clearest examples of how I like to work: close to the hardware, careful about interfaces, and interested in projects where software only makes sense because the electronics are doing their part too.