Skip to main content

Edge AI & embedded systems

Intelligence that fits in a microcontroller.

We bring machine learning models to STM32: sensors, audio and vibrations analysed on board, without the cloud, with millisecond latency and battery-level power consumption. From dataset to board, with STM32Cube.AI and ONNX.

In practice

Models tailored to the MCU

Compact architectures for audio, vibration and sensor signals: 8-bit quantization, pruning, feature selection (e.g. mel spectrograms) to fit in the available KB of RAM and flash.

ST toolchain

Training in PyTorch or TensorFlow, ONNX export, conversion and optimization with STM32Cube.AI; validation on the target, not just in simulation.

Energy budget

We design for battery life: duty cycling, event-driven inference, peripherals off when not needed. We measure real consumption.

From prototype to product

Firmware, field data collection, robustness tests against noise and real conditions, documentation for industrialization.

Stack and technologies

  • STM32 · STM32Cube.AI
  • ONNX
  • PyTorch
  • TensorFlow Lite Micro
  • Arm Cortex-M
  • Python

Project

Delivered

Detecting pipe leaks by listening to them

Acoustic anomaly detection on board an STM32: microphones listen to the pipes and a model recognizes the sound signature of a leak, with no invasive work on the plant — no holes, no downtime. STM32Cube.AI toolchain via ONNX, with extreme attention to power consumption. Prototype built and validated.

Accuracy
97%
Latency
milliseconds, on board
Toolchain
ONNX → STM32Cube.AI
Power
optimized for battery

How we work

Four steps, always the same

  1. 01

    Signal collection and annotation in real conditions

  2. 02

    Design and training of the compact model

  3. 03

    Quantization, ONNX export, porting with STM32Cube.AI

  4. 04

    Accuracy, latency and power measured on the target; iteration

Frequently asked questions

Which microcontrollers do you work with?

STM32 families supported by STM32Cube.AI (Arm Cortex-M4, M7, M33 cores and newer); on request we evaluate other MCUs with TensorFlow Lite Micro or dedicated runtimes.

Is the cloud required?

No: inference runs on the device. Connectivity is only needed if you want to send alerts or update the model remotely.

How accurate can such a small model be?

It depends on the problem and the data: in our leak-detection prototype we reached 97% accuracy at battery-level consumption. We always measure on the target.

Get started

Ready to transform your business?

Contact us for a free consultation and discover how we can help you reach your goals.