Ningyuan Zhang

1. Relevance Modeling in the Bing Search Product Ads:

• Own the relevance model stack for several product scenarios. Trained models such as MTDNN, TwinBERT, etc., with new features (product attributes, category features, etc.), completed offline evaluation, online A/B testing, and shipped models to production.
• Applied GPT and Mistral model-based AI annotation techniques (LLM prompt engineering) to detect irrelevant query-ad pairs and shipped to production.
• Owned offline data processing ETF pipelines, and online embedding publish pipelines on Azure.
• Owned the online model monitoring system, monitoring real-time feature coverage, embedding defects, and model metrics.
• Built automation tools to collect user feedback and process it into validation datasets.

2. End to end product engineering for Multimedia Ads:

• Built external, asset-based multimedia ad products end-to-end on Azure Data Factory and shipped to production.
• Built a monitoring system for MMA online serving and metrics tracking.

• Requisition Intelligence Project: managed project timeline and led scrum/user case meetings; built forecasting models on sparse time-series data to predict the purchase date for regularly purchased items and recommend quantity, supplier, ship location, etc.
• ActiveER Project: built a model application to cluster similar records and generate one standard record for each cluster based on a dedupe algorithm; built and tested APIs using FastAPI and pytest; built a UI using Plotly Dash, CSS, and HTML.
• Ticket Intelligence Project: Used NLP approaches to analyze the text content of tickets filed by customers, extracted and summarized the top frequent issues reflected by tickets, and helped the customer support team target the main issues to focus on in the Ariba Procurement Platform.

• Researched on financial time series data analysis and forecasting using deep learning
• Designed and implemented different DL architectures using Keras, such as CNN, attention-based LSTM and hybrid CNN-LSTM, to capture time series patterns
• Built an automatic machine learning pipeline with the research team, wrote Python modules for model stacking methods and time-series feature extraction using deep learning models