Work Experience Summary

4+ years of experience designing, training, and optimizing up to 519B scale large language models using distributed training techniques, contributing to a Top-3 ranking in the Korean World Best LLM project.
4+ years of experience partnering with cross-functional teams and clients to deliver AI solutions—including LLM fine-tuning, dataset synthesis, and evaluation system development—for projects valued at over $4 million.
4+ years of research experience in autonomous robotics and reinforcement learning, delivering 8 peer-reviewed publications and 2 patents through innovative solutions to real-world challenges. </aside>

WORK EXPERIENCE

Improve General Performance of Robot Policy Model: Developed a robot data diversification pipeline by training a multi-view Cosmos-Transfer 2.5 model to maintain consistency across different camera views, improving the OOD task success rate from 8.3% to 75% - https://aws.amazon.com/blogs/physical-ai/how-config-scales-robot-training-data-without-scaling-data-collection/
Open-source Release: Designed and implemented ESFT (Expert-Specialized Fine-Tuning for MoE) on AWS SageMaker - https://github.com/aws-samples/sample-ESFT

ICML tutorial workshop (2026): Presented about Physical AI data pipeline in the topic “From Digital Agents to Physical Intelligence: The Agentic Harness as a Unifying Architectural Pattern”
ICML Amazon Booth Demo (2026): Presented about robotic data diversification in the topic “Physical AI on AWS: Enabling robots to see, plan, and act in the real world”

Open-source Release: Main contributor to the development and release of SKT’s LLMs A.X-3.1, A.X-4.0, and A.X-K1 on Hugging Face, contributing to a Top-3 ranking in the Korean World Best LLM project
Large-Scale Distributed Training: Implemented multi-node training for 70B-scale Dense LLMs using FSDP, and 519B-scale MoE LLMs using Megatron-LM
Model Performance Improvements: Implemented Knowledge Distillation, improving general performance of 7B-scale LLM by 20%
Training Optimization: Developed "Preference Packing" technique for DPO that reduced training time and memory usage by 35% - arxiv.2602.24082v1
Training Workflow Management: Designed a post-training workflow to resolve issues in use cases and improve collaboration with the data generation team