Projects

AI Agent solution planning for manufacturing operations — field knowledge search, work guidance, and operator support. Ran a Qwen2-VL safety-monitoring PoC for CDQ dust discharge work, then shifted to YOLOv11 for deployable PPE monitoring under on-premise constraints.

• LLM
• AI Agent
• VLM
• Qwen2-VL
• YOLOv11

Developed and stabilized the learning pipeline for slab automatic rotation control from thermal camera data, combining BC pretraining, offline SAC, and world-model validation. Resolved SAC divergence and policy drift via conservative BC weighting, reward/target-Q clipping, and safety-violation filtering.

• Reinforcement Learning
• Offline SAC
• Behavior Cloning

U-Net segmentation for metal flow detection with angle-calculation logic using 25%/75% bondline points and horizontal reference lines. Refined labeling policy and completed additional angle labeling for model improvement.

• U-Net
• Segmentation
• Computer Vision

Expanded a 3-class to 5-class defect classifier, training on 403 images to 77.05% test accuracy under difficult label-boundary conditions. Applied split augmentation and proposed parallel learning of multiple preprocessing views.

• Image Classification
• Augmentation

Designed ROI, polygon, edge-detection, and line-connection logic for machine-room belt deviation detection. For the tail section, applied perspective warp and trained a segmentation model on ~300 labeled images to cut false positives.

• Computer Vision
• Segmentation
• CCTV

Tested multiple ML approaches for LDG holder level optimization and venting minimization, evaluating 12 models with 5-fold CV under limited segment-level data while balancing performance against compute constraints.

• Optimization
• ML
• Energy

Developed CCTV-based TP/Scrap outlet jam detection and normal/abnormal classification with limited abnormal samples. Improved the 9t Apex Point metric from x-coordinate output to a distance between the top/down midpoint and crossing point.

• Object Detection
• CCTV
• Classification

Developed time-series anomaly detection models for industrial equipment using AutoML and advanced ML, extending toward data observability and drift-resilient classification — published in IEEE Access.

• Anomaly Detection
• AutoML
• Time-Series