Lee DongWon

Hello! I’m a Software Convergence undergraduate at Kyung Hee University and an undergraduate researcher at the EMIL Lab, advised by Professor Sungha Choi. I study KV cache quantization to cut LLM memory for on-device inference, aiming for work that reaches real products. I also love AI tools and heavily use agentic AI like Grok Bot.

Experience

Undergraduate Research Intern, EMIL Lab, Kyung Hee University
Jul. 2026 – Nov. 2026
  • Evaluated an initial KV cache quantization approach for long-context LLM reasoning; a modified ThinKV implementation showed a 47.5 percentage-point accuracy gap versus FP16 on an AIME 2025 subset (10 problems × 4 responses, DeepSeek-R1-Distill-Qwen-14B).
  • Redirected research to importance-based global KV cache quantization, targeting lower memory use while preserving reasoning accuracy.
Robotics System Software Engineer Intern, ROBROS
Dec. 2025 – Feb. 2026 · Seoul, South Korea

Built and tested a humanoid SLAM image-streaming module on edge devices (Jetson Orin Nano, Intel NUC) over heterogeneous networks, tuning CycloneDDS and Zenoh in C++ and Python.

ROBROS camera streaming project →

Undergraduate Research Assistant, Kyung Hee University
Jul. 2025 – Dec. 2025 · Yongin, South Korea
  • Led end-to-end data analysis for a joint study with social science researchers on ChatGPT as an AI textbook alternative.
  • Processed multimodal data, built ML/DL models, and validated findings through empirical analysis.
Sergeant, Republic of Korea Army
Jul. 2022 – Jan. 2024

Completed mandatory military service.

Current Work

Memory-Efficient Long-Video Understanding (Capstone)

– Present

Streaming long-video understanding with Qwen2-VL-7B: on top of InfiniPot-V KV cache compression, I store the kept KV cache in INT4 instead of FP16, so the model needs fewer KV bytes at the same token budget.

View project →

Memory-Efficient KV Cache Quantization for Reasoning LLMs (EMIL Lab)

– Present

Studying memory-efficient KV cache quantization for reasoning LLMs (in progress).

View project →

Projects

  • 찰칵의 덕후 logo찰칵의 덕후 — Photo Studio & Photographer Finder App — 사진관과 사진가를 찾아 비교하고 예약·결제까지 하는 모바일 앱 창업 프로젝트 (Mar. 2024 – Jul. 2025). View project →
  • GPU-Accelerated LLM Inference (Mar. 2025 – Present) — GPU kernel optimization and self-hosted LLM serving with vLLM on limited GPU resources
  • ROBROS — Real-time Camera Streaming for a Humanoid Robot — Live multi-camera video over wired + Wi-Fi networks on the IGRIS-C humanoid, now part of its public SDK Camera API. View project →
  • Data-Efficient 3D Reconstruction — View selection via dataset distillation for 3D Gaussian Splatting (Mar. – Jun. 2026). View project →
  • LLM Inference Optimization on DGX Spark — Applied speculative decoding (Qwen2.5-0.5B draft), reducing TPOT 73–75% on batch-1 synthetic benchmarks (E2E −71% at 512-token input), with MMLU-Pro unchanged (42.1%, 1000-question subset) and HumanEval 83.5% vs 82.3% baseline. Evaluated and rejected uncalibrated FP8 KV-cache after its accuracy collapsed. View on GitHub →LLM inference optimization on DGX Spark summary
  • 3DGS Keyframe Distillation — Designed sharpness-and-coverage keyframe selection for 3D Gaussian Splatting (gsplat), evaluated on three TUM RGB-D sequences with two seeds and 7k iterations: 30/60-frame selections averaged +0.55 dB PSNR (range 0.0–1.5) and +0.015 SSIM (0.004–0.025) over uniform sampling across six scene/budget settings, using 1.7–21.1% of available training frames. Results varied by scene; desk60 PSNR was within seed noise, and blur-only selection outperformed the method on fr2_xyz. LPIPS not measured. View on GitHub →3DGS keyframe distillation summary