I am an M.S. student at BoanLab, Department of AI Convergence, Dankook University, working on AI infrastructure.
My work focuses on GPU workloads on Kubernetes: observing and quantifying GPU interference with eBPF under
time-slicing/MPS sharing, and runtime security for LLM serving (sidecar gateways, policy automation).
Training and serving large models ultimately comes down to how efficiently a handful of GPUs can be driven.
That is why I care more about what sits below the framework — schedulers, networking, memory, and kernels — than the framework itself.
Experience
Graduate Researcher (M.S.)
Mar 2025 – Present
BoanLab, Dankook University
GPU sharing interference: quantifying and attributing interference under time-slicing / MPS, with eBPF-based per-pod observability (gSonar)
GPU sharing platform: building a GPU sharing platform for interactive workloads on Kubernetes (gShare)
LLM serving security: a sidecar-based security gateway for LLM APIs (Sentra) and LLM-driven runtime security policy automation (KubeScribe)
Network performance analysis of mTLS in service meshes
#GPU#eBPF#Kubernetes#LLMSecurity
Undergraduate Researcher
Aug 2024 – Feb 2025
BoanLab, Dankook University
Introductory research on distributed training and GPUs
Performance analysis of parallelization techniques for large NLP models
#DistributedTraining#GPU
Undergraduate Researcher
Jul 2023 – Jul 2024
CIS Lab, Dankook University
Computer vision solution for the visually impaired
ML model training
Embedded board programming on Arduino
#ComputerVision#Embedded
Education
M.S. · Dept. of AI Convergence
Mar 2025 – Present
Graduate School, Dankook University
Lab: BoanLab
B.S. · Computer Science and Engineering
Mar 2019 – Feb 2025
Dankook University
Publications
Under Review
gSonar: Quantifying and Attributing GPU Interference under Time-Slicing and MPS
Jaeyoung Lee, Jaehyun Nam
IEEE Access·Journal
ckptGuard: 구조-인지 보호와 스트리밍 복원을 통한 LLM 추론 체크포인트 보안
ckptGuard: Securing LLM Inference Checkpoints with Structure-Aware Protection and Streaming Restore
Jaeyoung Lee, Jaehyun Nam
ACK 2026 (KIPS Fall Conference)·Domestic Conference·in Korean
Published
Time-Slicing GPU 공유 환경에서 eBPF 기반 파드 단위 관측 및 간섭 분석
eBPF-Based Per-Pod Observation and Interference Analysis in Time-Sliced GPU Sharing Environments
Jaeyoung Lee, Myeongdu Gim, Jaehyun Nam
KCC 2026 (Korea Computer Congress)·Domestic Conference·in Korean·Best Presentation Award
Sentra: A Kubernetes Sidecar-Based Multi-Layer Security Gateway for Protecting LLM APIs
Jaeyoung Lee, ChanUk Park, Jaehyun Nam
MobiSec 2025·Poster·KIISC Best Poster Award
KubeScribe: LLM-Driven Automation of Runtime Security Policies in Cloud-Native Environments
Jaeyoung Lee, Jaehyun Nam
ICTC 2025·Int'l Conference
쿠버네티스 기반 서비스 메시 환경에서의 mTLS 적용에 따른 네트워크 성능 분석
Network Performance Analysis of mTLS Adoption in Kubernetes-Based Service Mesh Environments
Jaeyoung Lee, Heesu Lee, Jaehyun Nam
ASK 2025 (KIPS Spring Conference)·Domestic Conference·in Korean
대규모 NLP 모델에서 병렬화 기법의 성능 분석
Performance Analysis of Parallelization Techniques for Large-Scale NLP Models
Jaeyoung Lee, Jaehyun Nam
ACK 2024 (KIPS Fall Conference)·Domestic Conference·in Korean
Awards
Best Presentation Award2026
KCC 2026 · Korean Institute of Information Scientists and Engineers (KIISE)
Best Poster Award2025
MobiSec 2025 · Korea Institute of Information Security & Cryptology (KIISC)
Sponsor Company Award (42Maru)2024
2024 SW-Centered University Digital Competition, SW Track · Software-Centered University Council
Finalist · Team 단컴한 인생 · Project Ghost Band
Excellence Award2024
2024 AI Idea Competition in All Fields · Bachelor's–Master's Linked ICT Core Talent Program, Dankook University
Team BAG · Project Voice Phishing Detection and Response Guide (GAN-based deepfake voice detection)
A self-hosted generative AI workspace for chat, reports, slides, images, and audio/video. Responsible for system design, deployment, and server operations.
Designed the split between the workspace (KloudChat) and the model/tool backend (KloudChat-LLM) behind a single gateway URL, so sign-in, history, and workspace stay up when the backend is down
Per-user LiteLLM virtual keys for per-person spend and audit logs, with monthly credit limits checked before every turn
Docker Compose deployment (web · API · PostgreSQL · PDF renderer) and server operations
