I build frontier LLM systems end to end: pretraining, post-training, GPU performance, and distributed inference.
I build frontier LLM systems end to end: pretraining, post-training, GPU performance, and distributed inference.
I work across the full LLM stack: large-scale pretraining, post-training and alignment, GPU-kernel optimization, and distributed inference. At AMD, I led MI455 Helios training optimization, created KernelForge—now core to AMD Hyperloom—and drove Day 0 inference bring-up for next-generation GPUs.
• Tech lead for large-scale training performance optimization on MI455 Helios.
• Designed and built KernelForge, a multi-agent GPU-kernel optimization system now core to AMD Hyperloom.
• Leading development of agentic systems for autonomous optimization across training and inference, including Day 0 LLM inference and performance bring-up for next-generation GPUs.
• Drove large-scale LLM pretraining optimization across Megatron-LM, PyTorch, JAX, and Primus.
• Built and optimized GPU clusters to support LLM training at 10,000-GPU scale.
• Advanced high-efficiency distributed inference for large models.
• Contributed across pretraining, alignment, and inference for Yi-Lightning, Yi-34B, Aria (an open-source multimodal MoE model), and Allegro (text-to-video generation).
• Optimized model training and inference on AMD and NVIDIA GPUs, and scaled large GPU clusters.
• Built Beago, an AI search app combining LLM fine-tuning and retrieval-augmented generation (RAG).
• Yi-Lightning ranked #6 overall on Chatbot Arena.
• Launched adaptive bitrate streaming (ABR) for ByteDance’s video-on-demand service.
• Led TikTok HTTP/3 performance work, increasing network throughput by 7% for billions of users.
• Designed CDN allocation algorithms that reduced costs by 10%+.
• Optimized PCDN architecture, contributing to 2%+ revenue growth.
Publications across ACM CoNEXT, ACM MMSys, IEEE INFOCOM, IEEE Transactions on Mobile Computing, IEEE Transactions on Wireless Communications, IEEE Transactions on Power Systems, IEEE Transactions on Sustainable Energy, IEEE Internet of Things Journal, and Applied Energy. Best Paper Awards: ACM CoNEXT 2018 — ABR Streaming of VBR-encoded Videos: Characterization, Challenges, and Solutions; IEEE PES General Meeting 2020 — Quantum-Secure Networked Microgrid.