Low-Power Edge AI Comes Alive with Edge Impulse and Qualcomm at Embedded World 2025

If you still think AI needs the cloud to be clever, Edge Impulse and Qualcomm are here to change your mind. At Embedded World 2025, this power duo took embedded engineers on a tour de force through the world of low-power edge AI—from anomaly detection on factory motors to on-device vision models that can spot dodgy coffee capsules before you sip your espresso.

Here’s how Edge Impulse is making machine learning more accessible, efficient, and downright practical for real-world applications.

Qualcomm + Edge Impulse: The Acquisition That Makes AI Easy

It’s official: Qualcomm has acquired Edge Impulse, and the message is loud and clear—AI at the edge is here to stay. Qualcomm brings cutting-edge silicon with high-performance NPUs; Edge Impulse brings the software stack that turns sensor data into deployable machine learning models without giving engineers grey hairs.

Their shared goal? To make AI easier to build, train, deploy, and optimise—especially when you’re dealing with limited data, tight power budgets, and tough environmental conditions.

Multi-Layered Low-Power Edge AI for Smarter Devices

Edge Impulse showed off some clever demo layering at the booth. In one standout showcase, engineers used:

• A lightweight object detection model to first detect whether a car (or coffee capsule!) is in the scene.
• A secondary vision language model (VLM) to extract features like vehicle brand or potential defects—only when needed.

This cascade architecture means the heavy-duty model only kicks in after a trigger—saving battery, compute cycles, and time. It’s the AI version of “speak when spoken to.”.

Tap, Train, Deploy: Simplicity in Action

Worried you need oceans of data? Think again. One demo used a Thingy:53 to capture just a few taps’ worth of acceleration data, then trained a vibration-based anomaly detection model. Within minutes, the system could detect unusual patterns on an industrial motor—perfect for predictive maintenance without the cloud overhead.

Using Edge Impulse’s dashboard, it was easy to collect time-series data, train a model, deploy it to the edge device, and monitor anomalies in real time. All with a few clicks.

From Smart Motors to Smarter Manufacturing

Want more excitement? How about real-time visual inspection for quality control? In another slick demo, a conveyor belt of coffee capsules was monitored using two edge models—one for identifying items, another for flagging anything malformed or out of place. This is next-level factory automation without the server racks.

Even better, these models ran on an MCU-grade chip from STMicroelectronics—proof that you don’t need datacentres to do clever things with machine learning.

Final Thoughts: Power, Precision, Practicality

This year’s Edge Impulse booth made it clear: low-power edge AI isn’t science fiction—it’s a practical, scalable solution ready for deployment across sectors. Whether you’re building a smarter home sensor, diagnosing a motor on the fritz, or inspecting products on the line, Edge Impulse gives you the tools to turn raw data into reliable action.

And the best part? Your device doesn’t have to stay awake all the time. It just needs to be smart enough to know when to listen.

So next time someone tells you AI lives in the cloud, just point them to the Thingy:53 quietly listening for a problem it hasn’t heard yet.