← All projects
AIArchived2023

The Nose Knows

A middle-school science-fair project

A low-cost electronic nose that tells coffee from air. Coffee was a stand-in: the real question was whether a cheap sensor and a trained model could pick out a mixture of compounds by smell.

The 3D-printed nose enclosure sitting on a mug of coffee
4gas channels: NO2, CO, ethanol, VOCs
120,000+data points per class, per the poster
100/100per class on the poster's test set

Why

It started from articles about dogs that can sniff out cancer, and Benjamin Cabé's open-source artificial nose. The long-term idea is training the same kind of nose on the compounds linked to lung cancer. That work hasn't been done.

How it works

The sensor reads four channels, the Wio Terminal records them, and the saved dataset trains a classifier. A fan clears the chamber between samples. With air, all four channels stay flat. With coffee, every channel reads higher and climbs slowly.

Air: flat.
Air: flat.
Coffee: higher and rising.
Coffee: higher and rising.
From sensor to prediction.
From sensor to prediction.
The parts.
The parts.

How far to trust it

The model got every test sample right, but the poster isn't consistent about how many times the test ran, and the original data and training code no longer exist, so the train/test split can't be checked. It's one coffee-vs-air setup working, not a general accuracy figure.