Speaker
Description
This talk explores the technical implementation of integrating GNU Radio into AI-native computing environments by encapsulating GNU Radio flowgraphs within an inference serving framework.
Specifically, GNU Radio flowgraphs are deployed within NVIDIA Triton Inference Server models, enabling DSP pipelines to be managed and orchestrated using the same infrastructure commonly used for machine learning inference. In this architecture, GNU Radio flowgraphs function as reusable RF signal processing components that can be reconfigured or replaced on-demand. Combined with neural network models and other Triton workloads, these components form end-to-end RF processing pipelines that integrate traditional DSP and neural network inference. We compare and contrast this approach with our previous work where AI models were embedded into GNU Radio as custom Python blocks, discussing tradeoffs in terms of reusability, reconfigurability, and overall performance.
We will describe how this capability has been integrated into Deepwave's AirStack Edge platform, where SDR processing pipelines can be deployed, configured, and controlled through a Model Context Protocol (MCP) interface. Rather than exposing low-level GNU Radio operations such as block creation and graph reconfiguration, AirStack Edge exposes workflow-level configuration and control interfaces through MCP, allowing agents to configure and operate SDR pipelines without modifying their underlying topology. This architecture enables LLM-based agents to discover, configure, and execute SDR workflows through natural-language interactions while maintaining a secure, well-defined execution environment.
Finally, we demonstrate a live, real-world implementation showcasing an autonomous agent using natural language to orchestrate a GNU Radio flowgraph on a remote sensor, perform complex RF AI inference, and report results back to a central platform.
| Talk Length | 30 Minutes |
|---|---|
| Acknowledge | Acknowledge In-Person |