Alex Rosito
Most approaches to running neural networks on microcontrollers rely on TensorFlow Lite — runtime, dependencies, overhead. This project takes a different path.
The pipeline
Train a 784-128-10 network on desktop using Hasaki — a C++ CLI tool that trains feedforward networks and exports standalone C headers. No Python, no framework, no runtime. The exported header contains the weights and a predict() function that compiles on any C99 toolchain.
The hardware
ESP32-C3 Super Mini. $3. No additional components required for the core demo. Optional: LED matrix display for visual output.
How it works
The ESP32 hosts a web server with an HTML5 canvas. Any device on the same network opens the page, draws a digit, and the chip runs inference locally in under a millisecond — no cloud, no external API.
Memory footprint (INT8, 784-128-10)
RAM: 40,108 B / 327,680 B — 12.2%
Flash: 1,034,610 B / 1,310,720 B — 78.9%
Flash includes the full sketch plus the 390 KB INT8 model header.
Training
Trained with Hasaki Pro — Adam optimizer, dropout 0.3, L2 0.0001, early stopping. Converged at epoch 1981. Final Val Loss: 0.058. Validation accuracy: ~98.2%.
INT8 quantization passes the mean error threshold — less than 5% degradation vs float.
The exported header
#include "mnist_int8.h"
float input[784];
float output[10];
predict(input, output);
// output[i] = confidence for digit i
No heap allocation. No dependencies. If it compiles for your target, it runs.