NetKit Labs
Software that belongs on the metal.
Libraries and applications for MCU, MPU, and edge AI — built for predictable latency, tiny footprints, and C/C++ APIs you can ship in firmware.
Projects
Libraries and applications that make serious embedded software easier to ship.
Real-time keyword spotting on ARM Cortex-M4F. Continuous 16 kHz microphone audio, CMSIS-DSP features, and int8 TensorFlow Lite Micro inference using Zephyr RTOS on a 64 MHz nRF52840.
View repository →Embedded-first multi-modal neural network inference for MCU, MPU, and NPU. Float32 and int8 paths, arena allocation, CMSIS-NN / XNNPACK backends, and peer benchmarks against TFLM and TF Lite.
View repository →Zero-overhead containers for bare-metal and RTOS targets. Shared cores behind C++ templates and a type-erased C API — no hidden mutexes, no RTTI, no virtual dispatch on the hot path.
View repository →FPGA NPU on AMD/Xilinx Zynq-7020 with an 8×8 int8 systolic GEMM array, AXI4-Lite control, and AXI DMA tile transfers. Python host software drives inference and speech-command experiments.
View repository →Approach
Designed for firmware constraints.
-
Predictable memory
Static arenas and explicit buffers. On MCU targets, no heap games — allocation stays under your control.
-
Dual-language APIs
Modern C++ where it earns its keep; clean C APIs where toolchain reality demands it. Same cores underneath.
-
Measured on hardware
Bring-up and benchmarks across real hardware — STM32 Cortex-M4F, Arm Cortex-A53, and host CPUs — not just simulations.