As the capabilities and performance of both AI models and edge devices continue to progress, it has become possible, and increasingly valuable, to perform inference closer to the point of data acquisition. Managing the deployment of models, as well as their ongoing evaluation and improvement, proves to be a complex task. It’s typically necessary to deliver inference results, and potentially the raw inputs provided to a model, back to the cloud for further action or analysis.

Golioth’s IoT platform offers integrated device management and data service capabilities, including over-the-air (OTA) updates, device logs, remote procedure calls (RPCs) and more. Golioth Pipelines provide robust and flexible data routing and processing, enabling transformation of streaming data before delivery to a wide variety of final destinations.

With Golioth’s recent launch of AI capabilities, data can also be easily ingested for model training or routed to a hosted inference platform. AI models can be managed and deployed via Golioth’s OTA service, providing visibility into which model each device in a fleet is running, and making rolling out new models seamless.

Golioth recently partnered with Foundries.io and the Qualcomm AI Hub team to demonstrate how integrating Golioth’s SDK (known for its use with cellular-connected MCUs) into embedded Linux applications makes it easy to utilize models from Qualcomm AI Hub on edge devices.

The Qualcomm RB3 Gen 2 development kit is a great candidate for showing off vision models, such as YOLO (“You Only Look Once”) object detection, with Golioth delivering the model to the device, then streaming inference results and raw image captures to the cloud. The development kit is based on the Qualcomm QCS6490 SoC, and includes a Qualcomm Kryo 670 CPU with 8 application processing cores, a Qualcomm Adreno 643L GPU, and a Hexagon DSP for accelerating AI workloads. Additionally, it boasts a low- and high-resolution camera, as well as audio peripherals, and an array of sensors.

Golioth + Foundries.io + Qualcomm AI Hub

The first step in deploying an AI application to the Qualcomm RB3 Gen2 Platform is registering the device with the Foundries.io platform. The platform’s GitOps workflow enables deploying OCI images as containers running on the Foundries Linux microPlatform (LmP) distribution. To help developers leverage the cameras and acceleration hardware on the Qualcomm RB3 Gen2, Qualcomm Technologies, Inc. provides the Qualcomm Intelligent Multimedia SDK (IM SDK), including a suite of GStreamer plugins. These plugins can be used to create a pipeline that processes media, runs inference using a supplied AI model, and overlays results on the final image.

Building a simple GStreamer pipeline with the YoloNAS model from Qualcomm AI Hub is straightforward. However, though this application could be deployed to the Qualcomm RB3 Gen2 via the Foundries.io platform, data from the application cannot be observed remotely, and modifying the model requires building and deploying an entirely new application.

Integrating the Golioth Firmware SDK means that we are instantly able to receive logging information, see what model is active on the device‌ and deliver inference results and image captures to the Golioth platform.

Once the data reaches the cloud, Golioth Pipelines are used to route data to multiple locations. For example, the inference results can be transformed into JSON payloads and Golioth integrated timeseries database. Image data, on the other hand, can be routed to a blob storage destination, where it can be later accessed for offline model training or fine-tuning.

When new models are released, Golioth updates the application, triggering a download of just the new model and labels. This ensures that devices can acquire the latest functionality quickly and efficiently, without having to replace the entire application.

Real world use cases for Golioth’s Data Plane

The flexible nature of Golioth’s data plane opens up a wide variety of real-world use-cases when combined with Qualcomm’s Technologies’ powerful hardware and AI models.

• Industrial Manufacturing: identify manufacturing defects, predict machinery failure, and create a safer working environment.
• Smart Cities and Buildings: optimize vehicle and human movement, monitor environmental levels such as air quality and weather conditions, and reduce carbon emissions.
• Agriculture: monitor crop health, develop efficient irrigation strategies, and improve overall yield.
• Energy, Waste, and Utilities: detect when waste receptacles are full, effectively plan for increases in energy consumption, and monitor changes in water quality.

Next Steps

You can learn more about how to leverage Golioth and Qualcomm AI Hub by checking out the documentation and latest model releases, or join the community via the Golioth forum and Qualcomm AI Hub Slack channel.

Check out FoundriesFactory^TM IoT software platform to get your free trial today. 

Excited to hear more updates? Join the community of like-minded developers to connect, get support and get access to our exclusive virtual events at Qualcomm Developer Discord.

What else is trending for Internet of Things:

Read how CyberLink ports FaceMe to Qualcomm Hexagon NPU for facial recognition on edge devices

How Capgemini uses Qualcomm Dragonwing portoflio to enhance  railway safety with Edge AI

Learn how to optimise your AI model for the Edge

Browse highlights from Embedded World 2025


Watch developer Build Along sessions:


MCP IoT Agent for Snapdragon X Elite and Rubik Pi

Using TensorFlow to accelerate models on Qualcomm IoT devices

Docker and Qualcomm Dragonwing RB3 Gen 2 x FoundriesFactory

AMA session with Qualcomm and Edge Impulse

AprilTag and Qualcomm RB3 Gen 2