Hi, I’m SeungYong Lee, an AI research engineer with a passion for building practical systems across computer vision, generative AI, and scalable infrastructure. As a founding member of an AI startup (NXN LABS), I’ve led full-stack development for production-level AI—from raw data collection and model design to large-scale training, deployment, and monitoring.
Educated in mathematics and electrical engineering at KAIST, I’ve built a strong foundation for tackling complex problems in AI. Through hands-on work across research labs, AI startups, and industry settings, I’ve developed and deployed systems that bridge state-of-the-art models with real-world applications.
My work spans the full stack—from designing novel model architectures and scaling multi-node training, to deploying optimized inference pipelines for AI services. Lately, I’ve been exploring test-time scaling strategies for high-quality, controllable, and efficient image generation and Multimodal AI for agent-based systems.
To continuously explore uncharted ideas and build AI-driven products and services that deliver real value to people’s lives.
I aim to bridge research and reality, turning technical innovation into meaningful, human-centered experiences.
Full-stack AI Engineer
Experienced in leading complete AI pipelines—from data engineering and model design to multi-node training and scalable deployment. Skilled in bridging cutting-edge research with practical infrastructure for interactive, generative AI systems.
Researcher
Specialized in computer vision and image generation, with hands-on experience developing transformer-based architectures, diffusion models, and medical imaging applications. Presented at CVPRW and preparing further academic work.
Startup Founder
As a founding member of NXN LABS, I’ve led the development of production-level AI services in the fashion and avatar generation domain. My work is deeply rooted in user impact, rapid iteration, and ownership across the entire product lifecycle.
Deep Learning
PyTorch, CUDA, Diffusers, Transformers, Hugging Face, Detectron2, OpenCV, Numpy, Timm, SMPL
MLOps
Docker, Multi-node HPC (H100/H200), Deepspeed, Config Management(Hydra/OmegaConf), Weights & Biases, Spot Recovery, Model Optimization(Quantization, Compilation)
Cloud & Infrastructure
AWS (S3, EC2, FSX, SageMaker), GCP, Azure, Naver Cloud, RunPod, FastAPI, Nginx, MongoDB, Gradio
Tools
ComfyUI, Git, GitHub, Notion, Jira, Confluence, Cursor, MCP
As a founding member of NXN LABS, I led the end-to-end development of a virtual try-on system for fashion.
- Built complete data pipeline: web crawling → pre-processing → human validation → DB structuring.
- Designed generative architecture using DiT + Flow Matching, tailored for fine-grained control.
- Trained on HPC clusters using DeepSpeed + FSDP, monitored with Weights & Biases.
- Evaluated with benchmark scenarios (user-level fidelity, inference time).
- Deployed full-stack: FastAPI backend, optimized inference pipeline, and user-facing UI.
A full-stack generative AI system, from raw data to real-time try-on experiences.
Built controllable and consistent virtual humans using 3D modeling and image generation.
- Used SMPL and HMR to map human mesh from 2D inputs.
- Controlled pose/viewpoint in Blender, rendering synthetic 3D conditions.
- Generated images using Stable Diffusion, with LoRA, ControlNet, Adapter extensions.
- Rapid prototyping with ComfyUI, streamlining model and prompt iterations.
Bridging geometry-based modeling and diffusion-based generation for controllable virtual avatars.
Contributed as a research scientist intern to improve cancer detection with transformer architectures.
- Introduced ViT-based backbones to replace CNNs, achieving significant accuracy improvements.
- Designed Multi-FoV Transformer for better contextual field integration.
- Migrated framework from heatmap detection to DETR-style regression, improving mF1 by +4.5.
- Built entire experimental pipeline: data loader, loss functions, evaluation metrics, visualization tools.
Demonstrated performance gains of vision transformers in medical image analysis with clinical impact.
- CVPR Workshop 2024 - Industry Talk on Commercial Level Image Generation for Fashion Domain, especially Virtual Try-On
- CVPR Workshop 2025 (Upcoming Submission) - Virtual Try-On & Off through single Transformer arXiv
KAIST — B.S. in Mathematics & Electrical Engineering
NYU CAS — Exchange Student, Fall 2023