Today a microcontroller costing a few euros can recognise a gesture, a keyword or an unusual noise, with no connection and a battery that lasts for months. That is TinyML.

Why AI on the microcontroller

• Minimal latency: the decision is made on the device, within milliseconds.
• Privacy: audio and measurements never leave the board; only the result is sent.
• Low power: transmitting an event costs far less energy than streaming data continuously.
• Works offline: ideal where the network is missing or unreliable.

The boards we use

The Arduino family offers boards for every stage: versions with built-in sensors (accelerometer, microphone, environmental sensors) for fast experiments, Pro boards such as Portenta and Nicla for professional and industrial products, and the newer ARM-based boards for low-cost prototypes. When more memory or built-in connectivity is needed we also use ESP32-family microcontrollers.

How a TinyML model is born

1. Data collection straight from the sensor that will be used in production, under real conditions.
2. Training a small neural network, designed from the start for a few kilobytes of memory.
3. Quantisation: weights move from floating point to 8-bit integers, cutting memory and compute with minimal loss of accuracy.
4. Field validation: we measure accuracy, false alarms, response time and power consumption on the real board, not just on a computer.

Typical use cases

• Keyword spotting for local voice commands.
• Vibration classification for predictive maintenance of motors and pumps.
• Gesture and motion recognition for wearables.
• Anomaly detection on environmental measurements.

Once the Arduino prototype is validated, the same model can move to a custom-designed board, with the cost and size of a finished product.