What we made
An electronic nose. A gas-sensor array wired to a microcontroller, feeding a classifier: it smells a thing — and names the thing.
The spec fits on an index card:
- Nose — an array of MQ-series gas sensors, the $2-a-piece commodity kind
- Brain — an Arduino UNO R3
- Mind — a classifier sitting on top of the raw sensor channels
- Built — in one hackathon weekend
- Status — alive, and by now slightly famous: it won the Cursor Physical AI Hackathon in Almaty this July, then spent two evenings being passed around the tables at Ohm's Cafe as Exhibit A
That's the whole physical AI loop — sense, decide, act — for less than a dinner out. Which is exactly why it makes the point better than any slide: the loop became a weekend project. That's the entire story of the hype.
What broke
Everything interesting happened between the sensors and the classifier.
- Warm-up. MQ sensors have heaters, and heaters need minutes, not milliseconds. The demo starts before the sensor does. You learn to plug it in first and introduce it second.
- Cross-sensitivity. To a cheap gas sensor, everything smells a little like everything. Coffee, alcohol, a marker pen — the channels all twitch. The classifier's real job is untangling that.
- Drift. Calibrate. Then calibrate again next week. Then again. A MQ-series array is less an instrument and more a pet that needs feeding.
The classifier — the part that sounds like the hard part — was the fast part. A frontier model writes that pipeline now. The plumbing is where the week went, and that ratio is the honest lesson of the whole build.
What we learned next
Sensors lie, and the lying is the curriculum. If you want in on physical AI, this is the on-ramp: buy the loop — a cheap board, one sensor, one output — close it end-to-end, ugly allowed, and give it a deadline. This build has a hackathon to thank for existing at all.