ReSort

The problem

Recycling’s weak link isn’t the technology — it’s the moment of sorting. Once waste goes into the wrong bin, fixing it downstream is expensive, and mostly it never gets fixed.

The approach

ReSort moves sorting to the source. A Raspberry Pi, a camera, and a YOLO model running on the device: it sees the object, classifies it, and points at the right bin.

Notable pieces

Inference at the edge. The model runs on the device, not in the cloud. No network, no latency, no images leaving the bin — and streaming a trash can’s camera feed to a server isn’t something anyone should want anyway.

Optimizing under constraint. The hard part was never training the model; it was running it on a Pi’s budget at a frame rate anyone would accept. Input resolution and model size were trimmed right down to the point where accuracy started to give.

Honest about being a prototype. This is a working proof, not a product. The code is open on GitHub.