Senior Embedded Software Engineer | C/C++ | Embedded Linux/RTOS | AI Inference & Performance Optimization
I develop embedded C/C++ software across embedded Linux, RTOS, and bare-metal systems, targeting Arm Cortex-M/A, ESP32, and RISC-V platforms. My background includes real-time software, Linux device drivers, DSP, multithreading, memory-efficient design, hardware/software integration, and performance optimization, including work on aerospace and industrial systems.
My recent work focuses on embedded AI inference engineering: architecture-specific INT8 kernels, multicore execution, compact Transformer/Conformer enablement and acceleration, and FPGA-based neural processing. I contribute to semiconductor-vendor-maintained open-source libraries and validate optimizations on physical microcontroller hardware.
- Espressif ESP-DL — PR #347 — Submitted benchmark-driven AUTO-mode dual-core scheduling for large batched GEMM on ESP32-S3 and ESP32-P4. Reuses the existing multicore execution path for eligible INT8, INT16, and W8A16 workloads while keeping smaller shapes on one core. Hardware benchmarks show up to 1.36× faster INT8 and 1.91× faster W8A16 GEMM execution; these are operator-level, not measured end-to-end model gains.
- Espressif ESP-DL — PR #346 — Submitted ESP32-S3 INT8 Attention query–key optimization using Xtensa hardware accumulation (QACC) for eligible symmetric INT8 tiles. Measured 1.74× faster full Attention execution with identical outputs in the tested hardware workload; unsupported shapes retain the existing scalar fallback.
- Espressif ESP-NN — PR #46 — Merged batched INT8 fully connected acceleration using optimized 1×1 convolution and multicore execution, with safe fallback behavior and API compatibility. Measured up to 4.0× on ESP32-S3 and 40.7× on ESP32-P4 versus the prior batched-FC paths.
- Espressif ESP-NN — Issue #54 — Proposed and prototyped accelerated INT8
BATCH_MATMULattention using the 1×1 convolution path with nonzero zero-point correction. In an experimental Espressif TFLite Micro runtime, integrated accelerated FC and attention paths plus GELU and rank-5 strided-slice operator support, enabling complete compact Transformer inference and measuring up to 3.8× faster end-to-end execution on embedded hardware. This integration was a prototype, not a merged runtime contribution. - Google TensorFlow Lite Micro (TFLM) & Arm CMSIS-NN — Upstream contributions in development, focused on embedded inference runtimes and optimized neural-network kernels.
- MemKit — Creator & Maintainer — Open-source embedded C++26 container library with C++/C23 APIs, comprising 14 C containers and 32 C++ utilities. Includes ring buffers, lock-free queues, maps, and pools with static/arena, heap, and virtual-memory backing for MCUs and embedded Linux. Earned 56 GitHub stars and 8 forks soon after release.
- KWSKit — Multithreaded Zephyr RTOS firmware on Cortex-M4F, with streaming audio, DSP, and measured hardware timing and memory usage.
- adxl345 — I²C accelerometer Linux kernel driver with interrupt-driven acquisition, C++ userspace library, automated tests, and hardware validation on Arm Cortex-A53.
- uartcom — Register-level STM32F446RE USART peripheral driver and bare-metal C++ firmware, with interrupt-driven I/O, custom startup/clock configuration, and automated tests.
- NetKit — Portable C++ FP32/INT8 neural-network inference runtime for MCUs and embedded Linux, integrating CMSIS-NN, ESP-NN, and XNNPACK, with selected INT8 kernels up to 10× faster than TensorFlow Lite Micro reference implementations.
- NPUKit — FPGA-based INT8 neural-network accelerator on Zynq-7020, featuring an 8×8 SystemVerilog systolic GEMM array, AXI DMA tiling, and hardware/software integration validated on physical hardware.

