Deepwave presented “Deploying GNU Radio in Agentic AI Systems” at this year’s GNU Radio conference (GRCon2026). This session covered how signal processing and modern AI execution converge at the edge. Deepwave also open-sourced the AirStack Edge Model Vault to further support the community in building composable RF AI workflows.
Decoupling DSP and AI Execution
Embedding AI models directly into monolithic GNU Radio Python blocks presents significant engineering challenges, including GPU memory management, synchronous blocking in work(), and rigid flowgraph boundaries.
AirStack Edge solves this by running GNU Radio flowgraphs as modular backend components inside an NVIDIA Triton Inference Server environment. In this architecture:
- GNU Radio handles streaming DSP: Custom source and sink blocks manage sample buffers, turning flowgraphs into isolated, callable processing units.
- Triton composes the pipeline: Named tensor interfaces connect DSP modules to GPU-accelerated AI models (such as PyTorch or TensorRT) and downstream processing.
- MCP enables Agentic Control: By exposing model capabilities via the Model Context Protocol (MCP), autonomous agents—such as Deepwave’s Radio Intelligence Agent (RIA)—can discover, configure, and task remote RF sensors on demand.
Click here for the full presentation.
What’s in the Repository?
The newly released AirStack Edge Model Vault provides the modular model structures, configurations, and environment configurations needed to deploy hybrid DSP and AI pipelines.
To demonstrate how these pieces fit together, a complete step-by-step ASR Ensemble Tutorial is available on Deepwave Docs. This tutorial walks through building an end-to-end Automatic Speech Recognition pipeline that links a GNU Radio NBFM demodulator flowgraph directly to a PyTorch speech recognition model.
Explore the Resources
- GitHub Model Vault: deepwavedigital/airstack-edge-model-vault
- ASR Ensemble Tutorial: AirStack Edge Tutorial
- AirStack Edge Documentation: AirStack Edge Overview
- AirStack Edge Product Information: Deepwave AirStack Edge