ML Systems & Inference Engineer — I build at the lowest layers of AI execution.
Bare-metal C++, zero-allocation memory discipline, and hardware-accelerated inference pipelines.
Currently transitioning from CPU SIMD optimization into GPU kernel engineering and LLM inference infrastructure.

Languages: C++17/C++20, C, Python
Hardware Acceleration: Intel AVX2, FMA, SIMD Vectorization, Cache-Aligned Memory Layouts
GPU (Active Learning): CUDA C++, Triton — porting existing SIMD kernels to GPU
Inference & ML: Graph Neural Networks, Physics-Informed Neural Networks, Transformer Architecture
Systems: Zero-Heap Allocation, RTOS Concepts, IPC, Real-Time Telemetry Parsing
Protocols: CCSDS Space Packet Protocol (spacecraft telemetry standard)

Current Focus

• CUDA kernel optimization
• Transformer inference internals
• Quantization and memory-efficient serving
• LLM runtime systems
• High-performance AI infrastructure

Zero-heap C inference engine for Graph Neural Networks on edge hardware.
4.10µs mean inference latency via manual 256-bit SIMD (AVX2/FMA) register acceleration.
Framework-free. Zero runtime heap fragmentation. Runs as a background microservice daemon
with real-time JSON pipeline output and integrated Maxwellian divergence validation.

C AVX2 FMA SIMD Inference Edge AI Zero-Allocation

Native C++ YOLO inference engine optimized for robotic telemetry applications.
52.9+ FPS sustained throughput. Contiguous memory design, zero-overhead execution path,
bypasses heavy frameworks entirely.

C++ Object Detection YOLO Real-Time Zero-Allocation Robotics

High-throughput C++17 demultiplexer for real-time parsing of CCSDS Space Packet Protocol streams.
CCSDS is the telemetry standard used by NASA, ESA, and JAXA in operational spacecraft.
Designed for deterministic low-latency processing on edge hardware under hard real-time constraints.

C++17 CCSDS Spacecraft Telemetry Space Real-Time Zero-Allocation

My background is in squeezing maximum performance out of constrained hardware at the CPU level —
zero-allocation design, manual vectorization, deterministic execution paths.

I'm now applying that same discipline to GPU kernel engineering and LLM inference infrastructure:
porting SIMD-optimized kernels to CUDA, studying transformer inference internals (FlashAttention,
PagedAttention, quantization), and working toward meaningful contributions to open-source
inference engines like llama.cpp.

The hardware changes. The discipline doesn't.

• Memory is a first-class constraint, not an afterthought
• Benchmark everything — intuition without numbers is opinion
• Deterministic execution over convenient abstractions
• Understand the hardware before you trust the framework

Open to collaboration, feedback on my projects, and opportunities in ML inference infrastructure.
Master's student in AI — actively building in public.