Mingze Du

Built an automated screen-capture tool using Java and the MediaProjection API; implemented a cross-ROM compatibility layer (MIUI, EMUI, ColorOS) to resolve per-manufacturer frame-loss issues; designed a collection pipeline across Reddit, Twitter, TikTok and Instagram, achieving 10x throughput versus manual collection and accumulating 500,000+ valid frames. Fine-tuned a pre-trained BERT text classifier with TensorFlow transfer learning, designing platform-specific feature extraction pipelines for Reddit and Twitter; applied data augmentation to address data scarcity, achieving F1 0.87 for automated detection and classification of unhealthy food advertisements targeting minors. Designed a three-stage ETL pipeline (gzip compression, chunked upload, async ingestion via AWS Lambda); introduced MD5 idempotency checks to prevent duplicate writes, lifting upload success rate from 70% to 98% and reducing S3 storage costs by 35%. Built a multi-dimensional data visualization dashboard in React an

Designed a Collaborative Filtering plus Content-Based hybrid recommendation system for 50,000+ MAU, combining user-profile tags with TF-IDF cosine similarity; added a cold-start fallback for new users, improving click-through rate by approximately 20% versus content-only filtering. Used MySQL EXPLAIN to locate N+1 query bottlenecks and missing composite indexes; restructured tables and added indexes; introduced batch queries in the Node.js layer to replace loop-per-record requests, reducing query latency by 40% and bringing the recommendation API P99 within SLA. Led the definition of RESTful API documentation and a unified error-code system, standardizing request/response data structures and cutting average front–back-end integration cycles from 3 days to 1 day. Implemented recommendation display components in React with lazy loading and skeleton screens; wrote Jest unit tests and API documentation; completed grey-release and regression verification with zero production incidents.

Built a Python (Requests + BeautifulSoup) scraping system for 20+ e-commerce platforms with per-platform DOM parsing strategies and rotating User-Agent pools to bypass anti-scraping measures, increasing platform coverage from 5 to 20+ and raising update frequency from weekly to daily. Cleaned and merged multi-source data with Pandas; Power BI dashboard identified two high-potential regional markets, with findings adopted by management in the quarterly expansion plan.