Originally published at: Golioth Edge Inference using Bluetooth and Qualcomm hardware is a Snap - The Golioth Developer Blog
A few weeks ago we were at Embedded World 2026 showing off a new demo that highlights the capabilities of Golioth Connectivity and also hints at a future where we are using Canonical tooling alongside existing capabilities. We showcased a Bluetooth node running a gesture detection program developed by EdgeImpulse on an nRF52840. The results of the detection are sent to the internet using Pouch, the Golioth innovation that encapsulates data to make it easy to transit to the cloud. This includes encrypting the resulting packet and adding headers that make that transit easier. The Bluetooth Gateway reference implementation that we normally showcase on a microcontroller-based gateway running Zephyr was ported to run as a Snap on Ubuntu. In this demo we showcased it on an iQ9, a very powerful edge computer that was also running an entirely different demo (showcasing how the Gateway can be added to just about anything). That was a lot of different words, so let’s break things down. Edge Inferencing on a Bluetooth node Edge Inference and gesture recognition is a standard demo from our friends at Edge Impulse. We were able to create tailored models using the accelerometer on the Tikk (the LIS2DH12) talking to the attached nRF52840. We hit the button, detect motion, and then the TinyML model spits out the likelihood that the motion is one of 4 options (Idle, Snake, Updown, Wave). We display the output on the attached LED matrix of the Tikk board. Now what? Well, the model has weights/probabilities of the different results, but those are just numbers. We can send those up to the Cloud in order to see them and then possibly take action as a…