Portfolio
비공개 포트폴리오입니다.
direcf.github.io / personal / portfolio
AI · Vision · Multimodal
Seungjun Lee
AI / Vision Research Engineer · SK Telecom
I'm a Vision & AI Research Engineer at SK Telecom, building computer-vision and multimodal AI systems that make it out of the lab and into real products. My focus is video understanding — human behavior recognition, real-time safety detection, and Agentic LLM/VLM pipelines. I've taken several of these systems from research prototype to nationwide commercial service, with 15 filed patents, a Minister's Award (ICT Award Korea 2025), and a 3rd-place finish at the CVPR AVA Challenge along the way.
Education
KAIST — Korea Advanced Institute of Science and Technology
M.S. in Electrical Engineering · RITL Lab (Adv. Jong-Hwan Kim)
Thesis: CAL-ODWC — Continual Adaptation Learning for Object Detection under Weather Changes
2020.09 – 2022.08
GPA 4.15 / 4.3
Hanyang University
B.S. in Electronic Engineering
2014.03 – 2020.08
GPA 4.04 / 4.5
Cum Laude
Experience
SK Telecom
Vision Engineer · AI R&D Center / Vision Algorithm Team
Jan 2023 – Present
CareVia — Developmental-Disability AI Care
2023.05 – 2026.05
- Built a multi-camera action-recognition system end-to-end and developed models & algorithms for 9 behavior classes:
pushkickpush-pulllie-downhit-headfalldownrunloiteringjump
- Shipped a real-time commercial pipeline with face recognition (FR) + body re-identification (ReID) on embedded AI Box.
- Developed a pose-estimation model trained on 3M+ images for behavior recognition, deployed on AI Box.
- Minister's Award — ICT Award Korea 2025; live commercial service across 12 welfare centers nationwide since 2023.
Ego-view AI — “SynapsEgo”
2025.06 – 2025.09
- Built a video semantic segmentation (event localization) model and LoRA-fine-tuned latest LLM/VLMs (Qwen3-MoE, Qwen2.5-VL) to surpass SOTA (Gemini-2.5-Pro) for a National Police Agency video-summary & report-generation project.
- Designed a Recursive Agentic LLM pipeline for video segmentation and synthetic narrative generation.
- Delivered a real-time SOP-based work-procedure recognition Agentic pipeline (Yonsei Nursing College PoC).
SK Hynix — Industrial Safety AI
2025.04 – 2025.06
- Designed algorithm pipelines for 4 industrial-safety AI functions (fall, collapse, fire/smoke, intrusion), cutting false alarms by 96% and meeting the business-unit target ahead of schedule.
- Result: early qualification for the SK Hynix main contract.
Awards & Challenges
🏅
Minister's Award — ICT Award Korea
Ministry of Science & ICT · for CareVia, a nationwide AI care service
2025
🥉
CVPR AVA Challenge — 3rd Prize
Accessibility Vision & Autonomy Challenge · Keypoint Track
2024
🥇
SKT AI Frontier (4th cohort) — Grand Prize (1st place)
In-house AI expert program · project: Vision Retrieval System
2024
Patents
15 filed domestic patents (co-inventor) across four areas — video understanding, visual perception, agentic multimodal AI, and industrial safety.
Video Understanding · 5
Multi-view video understanding, VLM-based action recognition, semantic video segmentation, and automated video-summary report generation.
Visual Perception · 4
Occlusion-robust pose estimation, object tracking, person re-identification, and action recognition & monitoring.
Agentic / Multimodal · 4
Multi-agent hierarchical memory, multimodal event extraction & verification, and spatiotemporal-graph event-cause analysis.
Industrial Safety · 2
Real-time safety video-analysis systems for hazardous-environment monitoring.
Certifications
☁️
AWS Certified Generative AI Developer – Professional
Amazon Web Services · AIP-C01
2026.09
Technical Skills
LLM / VLM Agentic Pipelines
ML / DL
Languages & Infra
Let's build something.
Seungjun Lee · Updated Sep 2026 · Full details available on request