When people talk about AI today, they usually mean chatbots and large language models. But a quieter and arguably more practical revolution is happening inside physical products: wearables, industrial sensors and smart cameras that run AI directly on a chip.
On September 23, Jean-Samuel Chenard, Founder and President of Motsai, joined Michel Dubois, Chief Scientist and Head of AI Lab, Innovobot Labs, to unpack what it really takes to embed AI in electronic systems and to get the performance a product needs. Jean-Samuel has spent more than 15 years designing hardware that combines telecom, battery power and embedded computing, and he holds a PhD from McGill.
Here are the key ideas from the session.
Three products, three definitions of performance
Jean-Samuel built the talk around three real products, each showing a different meaning of “AI performance”:
- A wearable for soccer players that estimates distance and training load from motion data, running on a processor that costs about two dollars.
- Industrial predictive maintenance sensors that track vibration, temperature and sound, designed to run for years on a small battery.
- Computer vision systems for transportation logistics that classify events and detect and track objects in real time.
The same underlying ideas apply to all three. What changes is what “performance” means.
Why put AI in an electronic product at all?
When the data is too complex to hand-code. Human motion varies from person to person, so conventional algorithms struggle. For the soccer wearable, models are trained on large GPU machines, then made to fit on a chip that costs a few dollars, because the bill of materials multiplies into the final product price.
When moving the data is the real cost. Most of a device’s energy goes to the radio for communications, not to computing. By analyzing data on the device, the predictive maintenance sensor cuts what it needs to transmit by about 200 times. The edge does not replace the cloud; it feeds it better data.
When decisions must be made in real time. For vision systems, a decision has to land within a fraction of a second. If the system drops frames, algorithms that worked in the lab start to break.
Accuracy is the wrong headline
One of the strongest messages of the session: evaluate your model against business metrics, not only precision and recall.
A false alarm and a missed problem have very different costs depending on the industry. Too many false alarms, and users stop trusting the system.
Michel shared an example from a robot sorting metal from waste: in the morning, a demand for high-purity copper means no false positives can be tolerated. Later in the day, with demand for lower purity, the trade-off changes. In this example, going from 95% to 96% accuracy matters far less than understanding what each error costs the business.
Training data has its limits too: vision systems have been fooled by people wearing T-shirts with patterns the model had never seen. Rich training data, including examples designed to break the model, is essential.
How you get performance, and keep it
Good enough is often the goal. A model that is 95% accurate and fits a sellable, profitable product beats a perfect one that never reaches the market.
The cheapest inference is the one you never run. A low-power motion sensor can wake the AI only when something happens, which saves a lot of energy.
Prove the gain early. In early production runs, store the raw data on the device alongside the AI’s predictions, so you can compare results and improve your models over time.
Beware the demo trap. A model that works on a laptop is not a fleet of devices in the field. Getting there takes coherent data capture, secure updates for firmware and models, and constant iteration, because sensors, mountings and clocks all drift.
Three takeaways
- Performance is a system-level property, not just a model metric. It is a business problem, and your product’s definition of performance may be very different from someone else’s.
- Master the IoT before you add the AI. Conventional pre-processing and simple heuristics are still highly relevant. If a product only needs on/off, like a basic thermostat, it may not need AI at all.
- Cutting corners doesn’t save the work. It changes who pays and when. Planning up front prevents costly recalls, complicated updates and the burden of supporting a legacy version 1 when a much better version 2 is ready.
Need help bringing AI into your product?
Innovobot Labs helps companies turn embedded AI ideas into products that work in the field, from sensing and low-power hardware to computer vision, firmware and cloud connectivity.
Let’s talk if you are:
- Adding AI to a wearable, sensor, camera or other connected device
- Trying to fit a model onto low-cost or battery-powered hardware
- Moving from a working demo to a product deployed at scale
- Not sure whether your product needs AI at all