Lightweight and various-shaped object detection model for embedded system- YOLOv5-based w/ customized modules to reduce 56% FLOPs- Develop iterative pruning to improve at least 94% inference speed and reduce 70% model size- Empower NPU capability by model compression w/ quantization using Tensorflow- Design and develop auto-labeling and augmentation pipeline, saving 80% intensive labor workPredictive health management (PHM) and anomaly detection- Runtime self-learning and inference structure w/ real-time performance (<10ms speed)- Develop unsupervised, domain-agnostic algorithms for pattern discovery w/ 90% correlation- Develop adaptive learning mechanism w/ 1D-CNN autoencoder to perform anomaly detection

Host of large-scaled defect recognition system of IC packaging frontend/backend process- Cross-team cooperation to successfully achieve 10+ defect items recognition capability- Design and develop YOLO+CV for wire bonding to achieve 5μm-level detection capability- Automated ML platform to keep track of model performance to maintain production yieldPerformance enhancements by reducing 7% false alarm through data processing flow optimization- Develop ROI-focused algorithms for die bonding and saw singulationOptimize business process to reduce 70% personnel operation by ETL workflow integration

· All-in-one plasma etching equipment by integration of processing and measuring system· Cross-functional cooperation to build feasible design in software and hardware· Design and develop route optimization model for plasma etching to reduce 74% surface error· Provide web-based processing strategy recommendation system using image encoding skill