rtx-rf-classify

An inference engine that classifies radio signals by modulation type. It targets embedded RISC-V, where a Python runtime or a general-purpose ML framework won’t fit, so the whole thing is C++23. It runs as a service over TCP, and CMake picks the compute backend at configure time: OpenBLAS, hand-written SIMD, or portable scalar.

I wrote the tensor layer and its expression templates. Chained element-wise operations compile into a single fused loop instead of allocating a temporary buffer at every step, which is where most of the speed comes from. The convolution, normalization, pooling, and dense primitives sit on top of that. FFTW3 goes behind a type-safe wrapper so the compiler enforces plan lifetimes and buffer alignment rather than leaving them to convention.

Built with ACM.Industry for Raytheon.