I build fast, scalable
ML systems from first principles.
I am an intensely curious ML Systems Engineer specializing in GPU kernel optimization, AI inference, and low-level frameworks. I learn complex tech stacks rapidly and love squeezing every drop of performance out of hardware.
Right Now
Currently deep-diving into the Qwen3.5 architecture. I am aggressively profiling bottlenecks using PyTorch Profiler, writing a Fused RMSNorm kernel in Triton, and implementing a Gated Delta Net in Tile_lang (FlashQLA) to maximize inference throughput.
The TL;DR
I don't just use APIs; I break them down to understand how they work under the hood. My engineering philosophy revolves around a mechanical understanding of compute.
⚙️ Low-Level Mastery
Fluent in C++, CUDA, and Go. From writing custom autograd engines to building OS-level container runtimes using Linux namespaces.
🔥 Hardware Extraction
Experienced in optimizing for strict hardware constraints, achieving 110%+ of cuBLAS performance on local consumer GPUs, and deploying advanced optimizations for architectures like the NVIDIA Blackwell B200.
🧠 Compiler & Inference Focus
Deeply experienced with PyTorch, CuTe, Triton, and model quantization (FP16/INT8) to drastically reduce inference latency in generative architectures (Denoising & Autoregressive).
Featured Engineering Highlights
GPU Kernel Optimization & CuTe
Ranked #40 globally in the GPU Mode group competition by directing a custom AI agent to generate a high-performance Group GEMM kernel leveraging the NVFP4 datatype for the Blackwell B200. I also actively write deep-dive technical blogs breaking down complex memory layouts and the CuTe DSL.
Building Frameworks from Scratch
Engineered Vibegrad, a custom automatic differentiation library built entirely from scratch, and Simple-LLM, a GPT-style autoregressive model. I build these to master the mechanical constraints of computational graphs and efficient decoding strategies.
Gen AI Inference Acceleration
Rebuilt Stable Diffusion 1.5 from first principles (UNet, VAE, CLIP) to profile denoising bottlenecks. By implementing Flash Attention-2 and precision quantization, I reduced inference latency by 85% (33s down to 4.7s).