Applied Researcher & High-Performance Software Engineer¶
12 years of experience bridging the gap between deep learning research and bare-metal execution. Specializing in computer vision optimization for edge hardware and low-latency C++ inference engines.
Model Optimization & Training¶
Custom architecture design, training and fine-tuning for specific business KPIs and accuracy targets under hardware constraints.
High-Performance Engineering¶
Systematic profiling and optimization of inference pipelines. Focusing on quantization strategies, performance-accuracy tradeoffs, and efficient C++ implementations for edge hardware.
The Outcome¶
Production-ready vision systems that don't compromise between speed and precision. Real-time performance on edge devices.
Featured Case Studies¶
Jetson Orin Nano 7 W vs 15 W on YOLO26
The silicon idle floor is identical in both modes (2.79 W measured) — 7 W mode only caps the upper bound. For YOLO26 at variants bigger than n, 7 W trades ~50% throughput for ~25% power. Lower watts ≠ better efficiency.
YOLO26 on Jetson Orin Nano 8GB vs Hailo-8L (and Hailo-8)
Head-to-head accelerator comparison at 15 W. Jetson FP16 beats Hailo zoo accuracy at every size with no quantization; Hailo-8 wins on small-model throughput; the Jetson INT8 path I tried was Pareto-dominated by FP16.
Quantization Noise Analysis: YOLO26 on Hailo-8L
Layer-by-layer SNR analysis explaining why YOLO26-M loses 16% accuracy after INT8 quantization despite having more parameters than the Small variant.
Porting YOLO26n to the Hailo-8L (85 FPS)
Achieving a 13.7x speedup on Raspberry Pi 5 by moving a "non-supported" detection head to a templated C++ post-processor.
Non-Invasive Quantification of Viability in Spheroids
Published in Frontiers in Bioengineering and Biotechnology (2026). Combining classical CV and deep learning to non-invasively predict cell viability.
Optimizing SRNN (5x Speedup)
Profiling and optimizing the Shuffling Recurrent Neural Network (SRNN) using Custom CUDA kernels and PyTorch to achieve 5x faster training.
Need to unblock your edge AI pipeline?¶
I help companies squeeze every possible FLOP out of their hardware while maintaining the accuracy their business goals require. Whether you're stuck on a compiler error or an accuracy gap, let’s solve it.