LLMs Have No Nose, So We Built One: an AI Nose on the Arduino UNO Q

k66 (Lana Chen) — Sun, 18 Oct 2026 20:22:21 +0000

This is a submission for the Hacktoberfest Open-Source AI Challenge: Side Hustle.

▶ Demo video (3:11): youtu.be/n88pRiuib3U
▶ Code (MIT): github.com/PGASlab/app
▶ Try it without hardware: live UNO Q simulation

What I Built

LLMs have no nose

AI can read, write, see and listen. Ask a chatbot whether the milk in your
fridge has gone off, though, and the honest answer is "I can't tell, I can't
smell it." Smell is the one sense we use every day to keep ourselves safe from
bad food, and AI simply doesn't have it.

So we built one.

Every sensor is a small nose

A nose works because different receptors respond to different molecules. PGAS
copies the idea with three cheap metal-oxide gas sensors, each tuned to a
different family of spoilage gases:

MQ-135 smells ammonia and sulfides, the "rotten egg" end of spoilage.
MQ-3 smells ethanol, which bacteria and yeasts make as food ferments.
MQ-2 smells a broad range of volatile organic compounds.

On their own, each is just a resistor that changes with the air around it.

Together, they make one big nose

Wired to an Arduino UNO Q, the three small noses become one. The UNO Q's
microcontroller samples all three every second. Its Linux side compares them
with clean air, decides whether the food is fine, and then acts: the LED
matrix on the board flashes a warning, an alert goes out, and your phone shows
the verdict. No cloud, no account, no internet. It all happens on a board that
runs from a power bank.

People once tested food with a silver spoon, because silver turns black when it
meets the sulfides of rotting food. PGAS is the modern silver spoon. We built it
for night-market stalls and small food businesses, where selling spoiled food
is a health risk and throwing away good food is waste.

The side hustle. Many night-market stalls in Taiwan are side hustles: one
person cooking after a day job, buying ingredients in the morning and selling
until late. There is no food lab, only a nose and a guess about whether the
milk for tonight's tea or yesterday's leftovers are still safe. One bad batch
can cost a small seller their customers. PGAS is a pocket-sized second opinion
they can carry to the stall, and it runs from the same power bank as their
phone.

Demo

Here is one real test, photo by photo, on a bottle of milk that had gone bad.

1. The parts

Everything PGAS needs:

Arduino UNO Q (bottom right): an STM32U585 microcontroller and a Qualcomm QRB2210 running Debian Linux on one board, with a built-in 8×13 LED matrix.
Three gas sensors: MQ-135 for ammonia and sulfides, MQ-3 (the orange one) for ethanol, MQ-2 for broad volatile compounds.
A breadboard with a 10 kΩ / 20 kΩ voltage divider for each sensor, because the UNO Q's analog pins take at most 3.3 V and the sensors output up to 5 V.
A USB-C power bank. No laptop is connected.
A phone on the same Wi-Fi. The board itself serves the PGAS page.

2. Control test: no food nearby

First we ran the full test with no food anywhere near the sensors, to make sure
PGAS does not raise false alarms. All three sensors stayed at their clean-air
level (rises of +0.03 V, +0.02 V and +0.00 V), and the board decided
FRESH in 0.6 ms. The LED matrix showed ✓.

3. The suspect: a bottle of milk past its best

Our test sample is a bottle of fresh milk that had been left to spoil. The
second photo shows there really is milk inside. We poured some into a small
container so one sensor at a time fits over it.

4. The guided test, one sensor at a time

Press Start test on the phone and the page walks you through it. Each sensor
gets 5 seconds: 1 s in clean air, 1 s to move it over the food, 3 s measuring.
A big countdown ring and the live reading show what is happening.

MQ-135 first. The sensor sits over the milk while the ring counts down.

Then MQ-3. When a sensor is done, the page tells you to take it out and get
the next one ready. Here MQ-135 and MQ-3 are ticked and MQ-2 is next.

Press the button when the next sensor is in clean air. PGAS never moves on
by itself, so you are never rushed.

MQ-2 last.

5. The verdict: do not eat, do not sell

PGAS says Food has gone bad. The MQ-3 (ethanol) rose +0.21 V above its
own clean-air reading, past the 0.20 V limit, so the board graded the milk
DO-NOT-SELL in 0.2 ms. MQ-135 rose +0.02 V and MQ-2 +0.05 V. Ethanol is what
bacteria and yeasts produce as milk ferments, so MQ-3 is the sensor we expected
to react most.

The pop-up is written for anyone: one big verdict and one bar per sensor. It
also lists what the agent did:

Sense: 3 gas sensors, rise above clean air
Decide: DO-NOT-SELL on the UNO Q, offline, in 0.2 ms
Act: LED matrix alarm (flashing ✕), alert queued (Telegram not set up yet)

6. For the curious: the live gauges

Scroll down the same page and every sensor has a speedometer dial, its voltage,
its normal range and a live chart. Over the limit, the card turns red and
blinks.

MQ-3 in clean air: 0.37 V, inside its normal range.

MQ-3 held over the spoiled milk: 0.70 V. That is almost double, and the card
turns red.

30 seconds after taking it away: 0.38 V, back to its clean-air value. The
sensor recovers quickly, so you can test the next food straight away.

MQ-3	Clean air	Over spoiled milk	30 s later
Module output	0.37 V	0.70 V	0.38 V

Code

▶ github.com/PGASlab/app (MIT): the Arduino sketch, the Linux app, the phone dashboard, tests that run in CI, and a browser simulation of the UNO Q.

How I Built It

How it works: sense → decide → act

Step	What happens	Where
Sense	MQ-135, MQ-3 and MQ-2 are read at 1 Hz with the 14-bit ADC, averaged 16×	STM32U585 MCU (Arduino sketch)
Decide	For each sensor: rise over its own clean-air reading. An ONNX model turns the three rises into P(spoiled), checked by a safety rule	QRB2210 Linux (Python, ONNX Runtime)
Act	LED matrix ✓ or flashing ✕, alert, verdict on the phone	MCU + Linux

The two halves of the UNO Q talk over the Arduino Bridge: the sketch calls
Bridge.notify("pgas_sample", …) once a second, and Linux sends LED-matrix
frames back with Bridge.call("draw", …). The phone page is served by the
official App Lab arduino:web_ui brick over Socket.IO, so the phone only needs
to be on the same Wi-Fi or hotspot.

Comparing each sensor with its own clean-air reading taken seconds earlier
is what makes a 5-second test work. MQ sensors drift with temperature, humidity
and warm-up, so absolute thresholds are unreliable, but a rise within a few
seconds is not.

The open-source AI: a model that runs on the board

The verdict now comes from a small machine-learning model that runs on the
UNO Q itself, with no cloud:

Model: logistic regression in scikit-learn, exported to ONNX and run by ONNX Runtime on the board's Linux side (pgas/bag_model.py). The model file is 411 bytes.
Input: the three rises (MQ-135, MQ-3, MQ-2). Output: the probability that the food is spoiled. Above 80% it is DO-NOT-SELL, between 50% and 80% it is SPOILING.
Training data: our own real measurements: the clean-air control (photo 2), the spoiled milk in the guided test (photo 9) and an earlier spoiled-milk test. Each one is repeated with ±0.01 V of noise, the spread we measured on a sensor at rest.
Safety net: three samples is a start, not a validated model. So the board also runs the simple rule (largest rise ≥ 0.10 V SPOILING, ≥ 0.20 V DO-NOT-SELL) and keeps whichever verdict is more severe. A food-safety tool should never be talked out of a warning by a model that has seen three examples.

The verdicts in photos 2 and 9 were made by the rule; we trained the model from
those same tests afterwards, and it gives the same answers (7% spoiled for the
clean air, 93% for the milk). Every new labelled test makes it better:
add the rises to SAMPLES and run python -m pgas.bag_model. The repo also has
a script to optimize models for the UNO Q with Qualcomm AI Hub (pgas/aihub.py).

Hardware

Part	UNO Q pin	Notes
MQ-135	A0	AO through a 10 kΩ / 20 kΩ divider (5 V → 3.33 V max at the pin)
MQ-3 (Waveshare)	A1	same divider
MQ-2	A2	same divider
Heaters	5 V	all grounds common
Power	USB-C	5 V power bank for portable use

What we learned on real hardware

The browser simulation got the logic right, but the real board taught us more:

Sketch libraries. The Bridge library needs MsgPack, DebugLog, ArxContainer and ArxTypeTraits listed in sketch.yaml. We copied the exact versions from an App Lab example on the board.
No LLM brick on our App Lab version, so the advice text is rule-based.
adb shell sets an Android TMPDIR that does not exist on the UNO Q, which broke arduino-app-cli. Our scripts reset it.
Sensors read lower than datasheets suggest. Our "sensor unplugged" check rejected a working MQ-3, so we lowered the floor to 0.25 % of full scale.
Plan for resets. During development the board sometimes rebooted, and the app did not always come back. We never fully pinned down the cause (loose wires on the breadboard are our best guess), so we added a keep-alive that restarts PGAS within a minute if it is not running.
UX matters even for a sensor project. Our first test took 4 minutes with a hood over the food. One sensor at a time, 5 seconds each, guided by the phone, it now takes about 15 seconds plus sensor swaps.

Why Does Open Innovation Matter?

A street vendor can't depend on a subscription, a cloud API or a phone signal.
Everything PGAS uses is open: Arduino App Lab and its bricks, the Arduino
sketch, Python, scikit-learn and ONNX Runtime for on-board inference, and this
code under MIT. Anyone can rebuild it for the cost of a board and three
sensors, retrain it on their own food, or swap in a better model, and no data
ever leaves the stall.

Prize Categories

Best Use of Arduino.

It runs a model on the board: an ONNX model decides with ONNX Runtime on the UNO Q, offline.
It is a physical agent that senses and acts: gas sensors in, LED-matrix alarm and alert out, with the decision made on the board in under a millisecond.
It uses what makes the UNO Q different: the MCU and Linux working together over the Bridge, the built-in LED matrix, and App Lab bricks for the web UI.
It runs offline on a power bank, so it can work at a market stall.

Limits

One spoiled sample and one clean-air control is a demo, not an accuracy study. The model has seen three real samples and the rule thresholds are provisional; both need many more foods.
MQ sensors need a few minutes of warm-up and drift with temperature and humidity.
PGAS is a reference tool, not a food-safety certification. If in doubt, do not eat it, and ask a professional.

Team

k66
Alice
29
Crystal
Cynthia
Built with Claude Code as AI pair programmer

🐾GitPet - No more lonely debuging

k66 (Lana Chen) — Mon, 16 Feb 2026 07:55:57 +0000

This is a submission for the GitHub Copilot CLI Challenge

What I Built

Do you ever feel lonely when you coding? GitPet is always there with you! When you git push, git commit, git ..., etc, GitPet then become stronger!

GitPet is a GitHub Copilot CLI extension, you can use prompt see my GitPet status.

Repository:

https://github.com/k66inthesky/GitPet

My Experience with GitHub Copilot CLI

I used GitHub Copilot CLI to create this project.

🐾GitPet - No more lonely debuging

k66 (Lana Chen) — Sun, 15 Feb 2026 17:32:31 +0000

Feed your pet with many Git Pushes, Commits, PRs, etc!

This is a submission for the GitHub Copilot CLI Challenge

What I build

Do you ever feel lonely when you coding? GitPet is always there with you! When you git push, git commit, git ..., etc, GitPet then become stronger!

GitPet is a GitHub Copilot CLI extension, you can use prompt see my GitPet status.
Author: k66
Repository: https://github.com/k66inthesky/GitPet

My Experience with GitHub Copilot CLI

I used GitHub Copilot CLI to create this project.

DEV Community: k66 (Lana Chen)

LLMs Have No Nose, So We Built One: an AI Nose on the Arduino UNO Q

What I Built

LLMs have no nose

Every sensor is a small nose

Together, they make one big nose

Demo

1. The parts

2. Control test: no food nearby

3. The suspect: a bottle of milk past its best

4. The guided test, one sensor at a time

5. The verdict: do not eat, do not sell

6. For the curious: the live gauges

Code

How I Built It

How it works: sense → decide → act

The open-source AI: a model that runs on the board

Hardware

What we learned on real hardware

Why Does Open Innovation Matter?

Prize Categories

Limits

Team

🐾GitPet - No more lonely debuging

What I Built

Repository:

My Experience with GitHub Copilot CLI

🐾GitPet - No more lonely debuging

What I build

My Experience with GitHub Copilot CLI