Jimin Bae jiminbae

👋

Undergraduate student at Kyungpook National University

Kyungpook National University — B.S. in Computer Science & Engineering (AI Computing Major)
- 2023.03 – Present
Undergraduate Intern @ HPC Lab, KNU
- Winter 2025 – Present
Visiting Student @ UC Berkeley Summer Sessions
- Summer 2026

Analyzed GPU performance for LLM fine-tuning workloads in an OpenStack-based private cloud
Compared GPU virtualization methods, including MIG and PCIe Passthrough, as well as scaling strategies such as Scale-up and Scale-out
Automated LLaMA 3.1 8B QLoRA benchmarking workloads using Axolotl and Docker
Monitored GPU utilization, memory bandwidth, and PCIe communication traffic using DCGM Exporter, Prometheus, and Grafana
Analyzed throughput differences across GPU allocation methods under equivalent VRAM conditions and communication overhead in FSDP2-based Scale-out configurations
Paper accepted at KCC 2026 (to appear)

Built a target-aware video face-swap pipeline that preserves selected target identities while replacing non-target faces
Combined YOLO face detection, BoT-SORT tracking, InsightFace/ArcFace identity matching, and InSwapper-based face replacement
Added quality-based fallback blur for small, occluded, or unstable faces
Generated annotated output videos with per-frame tracking logs and face metadata

Addressed the limitation of Spiking Neural Network-based time series forecasting models that underutilize inter-channel relationships
Quantified channel relationships in multivariate time series using an attention mechanism
Incorporated STE-based channel similarity information into spike encoding representations

Built an end-to-end pipeline that analyzes civil complaint content and recommends the responsible department
Achieved 78.64% zero-shot accuracy on Daegu Dalseo-gu civil complaint data using LLM embeddings and VectorDB-based similarity search
Reduced public officials' workload by detecting and blurring abusive or harassing content and blocking duplicate complaint histories

Built an AI-assisted grading and feedback prototype for teaching assistant and student workflows
Designed n8n automation workflows for assignment creation, submission intake, grade approval, and feedback regeneration
Integrated Supabase-backed data modeling with Next.js API routes for grading, queue management, and student result retrieval
Created interactive TA/student interfaces for rubric-based review, AI feedback editing, code/essay grading, and learning recommendations

Built an NVIDIA AGX Thor environment for running the Isaac-GR00T-N1 nut-pouring model
Designed and executed a SITL workflow in Isaac Sim
Deployed ROS on Jetson Thor and applied the workflow to real hardware through HITL

Grand Prize, Upstage AI Agent Hackathon — Team Rough and Ready (2025.10 – 2025.11)
Excellence Award, 2026 Google AI Agent Challenge — Team Forward (2026.04 – 2026.06)