AI that runs on the device, even without internet

There is a camera above the line, but someone still checks by eye whether the label is straight. Or you have sensors in the field that send every reading to the cloud, where a model forms an opinion seconds later, if the connection holds. Edge AI runs the model on the device itself: the decision is made on the spot, in milliseconds, and the images never leave the building.

When the cloud is an unnecessary detour

For many applications, the cloud is a detour. A machine that has to stop because a hand comes near cannot wait for a server in Frankfurt to answer. A camera in a greenhouse on a 4G connection cannot upload a photo every second. And a healthcare organization does not want images of residents to leave the building at all. In all those cases, it makes more sense to run the model where the data is created: on a small computer next to the sensor, inside the enclosure of the device, or in the PLC cabinet on the factory floor.

That saves more than response time. No stream of raw data goes outside, so you need no expensive connection and no monthly bill for computing power. The device simply keeps working when the internet drops. And you never have to explain to anyone where the images go, because they go nowhere.

What fits on a small device

A language model like the one in a chatbot does not fit on a microcontroller, and it does not need to. Most questions on the work floor are small and sharp: is this product good or bad, is someone in the danger zone, does this pump sound different from yesterday, how many crates are on the pallet. For questions like that, small specialized models are enough, and they run fine on a Raspberry Pi, a Jetson module, or a microcontroller.

We choose the hardware to fit the question, not the other way around. Sometimes a module with a small GPU is needed because you have to process high-resolution images. Sometimes a chip that runs a year on a battery and only wakes up when a vibration is off is enough. For a new device, we design the circuit board together with our PCB and electronics design service, so computing power, sensor, and power supply fit together. If it really does not fit on the device, we tell you right away and look at a middle ground: a small box on site that serves several sensors.

From examples to a working prototype

We start with what you already have. A folder of photos of good and bad products, a series of readings from a machine that failed once, sound recordings of a compressor. We train a first model on that. Within ten business days, that model runs on a development board next to your machine or sensor, and you see whether it picks up the right things. Often it turns out you need different examples than you thought, or that the lighting matters more than the model. That is why we test on site early.

From that prototype, we make it robust. The model is made smaller and faster for the chosen chip. The firmware around it makes sure the device starts up cleanly, updates itself, and sends an alert when something is off. We build that with Embedded firmware, so the model and the rest of the device are one whole and not two separate projects.

Your data and model remain yours

Everything the device sees or hears stays on the device. Only the result goes out: approved, rejected, a count, an alarm. If you want to collect those results to see trends, they go encrypted to a server in the EU and connect to your existing reporting. For the formal side, especially when people are in the picture, we build the rules of the GDPR and the EU AI Act into the design from the start.

The model, the training data, and the code are yours. You get them in a form you can retrain yourself when your product line changes, with documentation of how we built it and why. No subscription per device, no vendor who can switch the model off remotely. You pay for the build, and after that the device belongs to your organization.