Ya Chu Chang

🚀 Data Science and Data Engineering Specialist (PySpark, Databricks, SQL, Anomaly Detection, Cross-functional Collaboration) • Designed and implemented a real-time analytics system to ingest and process product performance metrics, enabling immediate monitoring, anomaly detection, issue diagnosis, and performance optimization, which reduced incident response time by 60% and supported 9 product teams across 4 regions. • Collaborated with cross-functional and internationally distributed external teams, including partners in Canada, to support their data analytics initiatives. Led data-driven discussions to design and implement product ring deployment strategies, and regularly presented deliverables to ensure alignment with stakeholders. • Proactively conducted statistical analysis and regression-based anomaly detection on product performance metrics to identify CPU spikes and memory leaks. Automatically prioritized issue severity and routed diagnostics to the appropriate developer teams, enabling real-time resolution and supporting leadership decision-making. The system achieved an 8.3% early detection rate, preventing rollback incidents and enhancing release stability. • Contributed to constructing and managing efficient 𝘀𝘁𝗿𝗲𝗮𝗺 and 𝗯𝗮𝘁𝗰𝗵 data pipelines on 𝗗𝗮𝘁𝗮𝗯𝗿𝗶𝗰𝗸𝘀 with 𝗣𝘆𝗦𝗽𝗮𝗿𝗸 and 𝗦𝗤𝗟, ensuring high data quality, integrity, cost-efficiency, and reliability across multi-region environments. • Designed and implemented a data pipeline solution that reduced 𝗰𝗹𝗼𝘂𝗱 𝗰𝗼𝘀𝘁𝘀 by 𝟮𝟬%, optimized 𝘀𝘁𝗼𝗿𝗮𝗴𝗲 by retaining 𝟴𝟬% of data volume while ensuring data availability, and met business requirements.

🚀 Model Development and Optimization: (Python, Pytorch, CNN, GPU, CUDA, Object Detection(YOLO), OCR(CRNN/Attention), Classification) • Developed an OCR pipeline using object detection and a Transformer-based self-attention model on GPUs, achieving 90% IoU in document detection and 99%+ accuracy with 95%+ coverage. The solution accelerated the credit card application process, saving 5-6 working hours per day. • Applied model pruning and MAC reduction techniques to optimize CNNs for deployment, reducing size by 68% and compute by 82% while preserving accuracy, supporting AI application scalability. • Optimized PyTorch data loading pipelines using parallel processing, enabling stable and fast model training and inference under system constraints. 🚀 Machine Learning: • Developed a classification model to predict 1-year credit default probability (AUC 0.80), using Explainable AI (SHAP) for model interpretability. Automated reporting of individual risk scores, percentile ranks, and key contributing factors to take data-driven action based on user-level predictions. • Developed a personalized fund recommendation system using embedding-based Collaborative Filtering to compute user–item similarity and deliver personalized investment suggestions. Achieved a 6× improvement in precision@3 over the baseline, validated through A/B testing. 🚀 MLOps: • Designed an MLOps framework with MLflow in PyTorch to automate data preprocessing and annotation, enable storage-efficient model tracking, and streamline end-to-end training pipelines, reducing model development time by 80% and improving data cleanup automation by 50%.