Building Ultra-lite Crash Detection on the Edge

This is an example post. Replace it with your own writing.

The Problem

Running ML models on microcontrollers is hard. Memory is tight, compute is slow, and the real world is noisy.

The Approach

We pruned and quantized a small CNN until it fit in under 10kb of flash memory, while maintaining >90% accuracy on our test set.

Results

The model ran at 50ms per inference on a Cortex-M4 — fast enough for real-time crash detection.