The Edge AI Air Detective

• Senses PM2.5, PM10, CO, CO₂, NO₂, VOCs, temperature and humidity, and runs all inference locally with no cloud round trip
• Forecasts the next few hours of hyperlocal air quality from the node's own sensor history using a small on-device time series model
• Identifies the likely pollution source from sensor signatures — traffic, garbage burning, construction dust, cooking smoke, crop residue — using sensor fusion, optionally with camera or microphone input
• Converts readings into an action the user can take that day: change a walking route, time the opening of windows, run an exhaust fan, move an outdoor activity indoors
• Keeps working through power cuts and network outages, storing data locally and syncing only compressed summaries when a link appears
• Scales to multiple nodes across a campus, colony or ward so a street-level pollution map can be built with zero per-node cloud cost

Mandatory hardware requirement: The solution must be built on the Arduino UNO Q. All AI inference must run on the board itself. Cloud APIs and external AI services may not be used for core decision-making. Sensors, actuators, displays and communication modules are your choice, but the UNO Q must be the computing platform of the submitted prototype.

Technical feasibility (the cheat sheet): Board — UNO Q pairs a Qualcomm Dragonwing QRB2210 quad-core Cortex-A53 running Debian Linux with an STM32U585 MCU for real-time sensing, giving both AI capability and deterministic sensor timing on one board. Sensors — PMS5003 or SPS30 for particulates, MQ135 and MiCS-6814 for gases, SCD41 for CO₂, BME688 or DHT22 for temperature, humidity and VOC index, Modulino nodes over I2C for fast prototyping. Software — Arduino App Lab with its Python and C++ bricks; the Linux side runs the model, the MCU side handles acquisition and actuation. Models — quantised TensorFlow Lite or ONNX Runtime, gradient boosting for classification, a compact LSTM or TCN for forecasting, optionally a small language model fully on-device for the advisory layer. Connectivity — on-board Wi-Fi and Bluetooth, LoRa or GSM as an add-on for outdoor nodes. Power — USB-C, or a solar panel with a LiFePO4 pack for a fully off-grid node.

For the Grand Finale, your build should include:

• A working UNO Q node that senses the required parameters and runs classification, forecasting and advisory entirely on-device
• The on-device forecasting model plus source-attribution logic, with training approach and quantisation documented
• A proof-of-offline run: pull the network, pull the mains, show it still sensing, deciding and advising
• A user-facing output layer that gives a plain-language action, not just a number
• A multi-node scaling plan with at least a two-node street-level view demonstrated, plus a bill of materials and per-node cost

The Grand Finale is judged on:

• On-device intelligence and offline integrity — 30%
• Sensing accuracy and quality of hyperlocal forecasting — 25%
• Source attribution and real-world actionability of advice — 20%
• Hardware craft, power efficiency and multi-node scalability — 15%
• Demo quality and clarity of storytelling — 10%