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 work on making LLM and video-LLM inference use less memory, mainly by compressing and quantizing the KV cache.
Studied memory-efficient KV cache quantization for reasoning LLMs (in progress).
Built multi-camera streaming for a humanoid robot SDK across a heterogeneous LAN/WiFi network (Python, partly C++), with SLAM in mind; validated on the physical robot.
Built a multimodal classroom dataset of ChatGPT use.
Led a 5-member team; shipped an iOS app.
Military service.
On top of InfiniPot-V compression (6K KV budget), I store the kept KV cache in INT4 instead of FP16. With Qwen2-VL-7B: Video-MME 63.5 vs 62.8 and MLVU 66.1 vs 65.8 (FP16), at ~1/3.1 the KV bytes (164 B vs 512 B per token per head).
In progress.