Edge AI means running artificial intelligence models where the data is born: in the factory, in the shop, on a vehicle, in a field. Less latency, less bandwidth, more control over data.

When the edge makes sense

• When the answer must arrive in real time, as in quality control on a production line.
• When there is too much data to send, such as video streams from several cameras.
• When data is sensitive and must not leave the company.
• When the connection is not guaranteed.

Hardware: from board to compact server

We choose hardware based on the real workload. A smart sensor only needs a microcontroller; for computer vision on one or two cameras we use Raspberry Pi-class boards with dedicated neural accelerators; for several parallel video streams or larger models we use modules with an integrated GPU such as the NVIDIA Jetson family, or industrial PCs with NPUs. We always evaluate power consumption, heat dissipation and the operating temperatures of the environment.

Optimised models

A model designed for the cloud rarely runs well on an edge board. We adapt it with techniques such as quantisation, weight pruning and distillation, and export it to portable formats such as ONNX to run it with the most efficient runtime for that hardware.

Solutions we build

• Visual quality control: surface defects, labels, presence and position of components.
• Workplace safety: personal protective equipment, hazardous areas, people counting.
• Retail and public spaces: people flows and space occupancy, without storing personal images.
• Maintenance: analysis of vibrations, sounds and thermal images to anticipate failures.

Managing the system over time

An edge system lives for years: we plan remote updates of models and software, device health monitoring and collection of uncertain cases to improve the models with real data.