Mar 2023 – Present · 3 yrs · Seoul, South Korea · On-site

• Enhanced student model performance by 30% by leveraging Knowledge Distillation from teacher over 2T+ tokens; Reduced compute by 5x by discovering caching mechanism for teacher logits; Model deployed in Galaxy AI of Samsung S25
• Led large-scale (37B) post-training framework implementation supporting packing, context-parallel, and on-policy RLHF/GRPO
• Coordinated multiple cross-division teams of 12+ engineers for on-device State-Space-Model (SSMs), establishing weekly timelines, performance audits, and driving KPI achievements of quantization, LoRA, and long-context metrics, with 3x throughput increase
• Achieved 70% faster convergence in LLM Upcyling and Matformer model-nesting via principled init and custom importance metrics
• Saved millions of dollars by leading a feasibility analysis of TPU v5-lite for LLM training; identified critical VRAM limitation, library support gaps (e.g., flash-attention), and unstable platform libraries which led to the rescission of a large-scale contract

Mar 2017 – Mar 2023 · 6 yrs · Seoul, South Korea · On-site

• Spearheaded resolution of persistent gradient explosions in large (70B) model training by identifying fundamental architectural/init issues in Pre-LN models. Enabled stable deep-thin (1000+ layer) transformers, validated across NLP, Vision and Speech teams
• Conceived and managed a team for improving dialogue response selection, slot-filling, summarization via test-time compute
• Reduced Question Answering errors by 40% by proposing and leading the use of LLM-generated synthetic data-augmentation
• Achieved state-of-the-art performance on multiple competitive Question-Answering leaderboards (SQuAD, MS-Marco, HotpotQA, TriviaQA, NQ), using Retrieval-Augmented-Generation (RAG); model deployed on 50M+ Samsung smartphones, support chatbots
• Upgraded Samsung.com’s rule-based support chatbot with multi-task transfer-learning ML, improving accuracy from 70% to 96%

Jan 2016 – Mar 2017 · 1 yr 2 mos · Seoul, South Korea · On-site

• Reduced delivery time for Artik Cloud, IoT.js SDKs from months to hours by contributing OpenAPI SDK-generator to Swagger-Codegen
• Developed contextual framework to automate actions on Samsung Tizen using event triggers, deployed on millions of Samsung TVs

May 2014 – Jul 2014 · 2 mos · Suwon, South Korea

• Created smartwatch apps for summarization, spam-detection and event extraction