Amidst the LLM craze, does anyone still care about old machine learning algorithms?

I've built my own framework that allows embedding, quantization, and self-retraining on microcontrollers using C++ from scratch, currently mainly for tree-based model families (like Random Forest, xgboost...). It can compress and train the entire MNIST dataset of 70,000 images on ESP32 with only 3MB of RAM while still achieving an accuracy of up to \~94% across 10 classes (models size about 600 KB of RAM). This is intended to help the model adapt without having to reload the code into the microcontroller.

Everything is here, including source code, demo, and documentation: https://github.com/viettran-edgeAI/MCU

Although it's designed to handle tabular data, I chose to demo it with a simple computer vision application for visualization.

I spent a lot of time on this project, it didn't rely heavily on AI, and I can explain every line of code. I'm open to discussing anything. I hope everyone can provide some feedback or suggestions. In my country, it seems like now they only care about LLMs; every paper tries to cram LLMs in and they don’t care about these older algorithms anymore—they just brush them aside.

https://redd.it/1skdyr7
@r_cpp