About

Promoted from Senior Software Development Engineer. Distributed training, LLM inference, and GPU kernel performance.

• MXFP8 training for DeepSeek-V3: delivered the full MoE training path in TorchTitan/TorchAO with a FlyDSL backend on MI355X, now at 1.2× BF16 end-to-end. Designed a ragged MXFP8 wgrad grouped GEMM reaching 2.27 PFLOP/s (45% of peak), 13× the prior Triton kernel and up to 3× BF16 hipBLASLt; being upstreamed to PyTorch Inductor.
• Large-scale MoE pre-training: scaled DeepSeek-V3 671B to 1,024 MI325X GPUs at 96% scaling efficiency with Expert Parallelism and FP8 grouped GEMM kernels; worked with the AMD and Meta PyTorch teams on TorchTitan/Primus-Turbo and co-authored published results showing a 2.77× end-to-end training speedup.
• MXFP8/MXFP4 kernels: built CDNA4 forward and dgrad grouped GEMMs with XOR-swizzled LDS layouts, ping-pong software pipelining, XCD swizzle for L2 locality, fused quantization epilogues, and Split-K for K-heavy shapes, gated by bit-exact operator parity tests.
• RL post-training on ROCm: led end-to-end enablement of vime on MI355X across Megatron training and vLLM rollout, validated the GRPO pipeline, upstreamed ROCm fixes, shipped a prebuilt container, and published the work on the vLLM blog.
• Inference: delivered launch-day ROCm support for OpenAI gpt-oss-120B/20B on MI300X/MI355X, and built TTFT, throughput, concurrency, and Kineto/Perfetto profiling workflows plus MoE prefill/decode roofline studies that guided serving and kernel decisions.
• Recognition: Next 5% Award presented by AMD's CEO, plus 2 Executive Spotlight and 6 Spotlight awards in my first year.

Fine-tuned Code Llama 13B into a text-to-Cypher model for natural-language graph querying, and built semantic search over a Neo4j graph of about 3 million research papers.

TA for ECE-GY 6143 Machine Learning — answered student questions, guided assignments, and ran regular office hours and review sessions.

Built educational materials for ML system deployment on NSF-funded cloud testbeds, covering load balancing and scaling with Kubernetes. Assisted Prof. Fraida Fund on the "Fount" project.

Implemented few-shot intent classification with BERT to demonstrate the impact of synonym-based text augmentation, and built educational materials on the role of complete methodology reporting in reproducibility — incorporated into the UCSC curriculum.

Built Python automation frameworks for end-to-end testing across 12+ peer groups, cutting functional test time by 80% through scripted failure-case simulation. Also built a pre-check build tool that cut missing-system-constant failure identification from 1.5 hours to 30 seconds.

Built a deep-learning model with FastText embeddings to predict financial risk in textual statements, highlighting potential risk passages in documents.

Coursework across Machine Learning, Deep Learning, Cloud Computing, Big Data, Internet Architecture & Protocols, and Computing Systems & Architecture.

Bachelor of Technology, Information Technology — where I first picked up Python programming and machine learning.

A UNet diffusion-model training framework in C++/CUDA with HIP support for unconditional diffusion training and inference on NVIDIA and AMD GPUs — reaching ~40% of PyTorch (torch.compile) end-to-end training speed.

Fine-tuned T5 for news-article summarization (ROUGE-L 0.42), packaged in a Docker container and served via a Flask web app.

Contributed to Andrej Karpathy's llm.c, making the CUDA kernels portable to HIP to add support for AMD devices.