Yu Shao

• Developed a real-time intent detection (RIDE) model for different customer service channels, including chat and call centers. The unlabeled data were clustered and labeled by GPT-4; the data were then used to train on a pre-trained BERT model, achieving a 90% accuracy/F1 score; the model was subsequently registered and deployed as a real-time service. Currently, I am working with our two-pizza team to further improve the model, adapting to the business need. The end goal is to integrate the model into a service powered by LLM, so that minimum effort from live agents is required.

• Used deep Q learning/reinfocement learning to develop a customer personal experience recommender called next best action or NBA. The model recommended best coverage and action item to customers based on different customer types and customer journey stages. Deployed the model through Azure Machine Learning. Conducted A/B testing to evaluate model's performance. Projected net profit reaches $136,300,000.
• Built a named entity recognition model based on the large language model BERT. The training data was synthesized by ChatGPT. Achieved an average accuracy of 85% and f1 score of 86%. That model was targeting to be used as part of the new Geico chatbot and customer service platform.

• Utilized various statistical models to help the company determine its conversion, churn and upgrade triggers as well as the North Star Metric.
• Utilized XGBoost to predict users churn between a 4-8 weeks window with accuracy at 83%.
• Utilized a machine learning model (CART) to predict user upgrade to our elite tier with accuracy at 95%.
• Performed the cohort analysis to help the leadership team better understand the churn rate change
• Developed and maintained ten dashboards to help the leadership team keep close monitoring over the product, marketing, financial and customer service statuses.

• Developed machine learning models (e.g. provider fraud, inter-departmental record linkage, Covid contact tracing) through cloud-based infrastructure to facilitate continuous integration/continuous deployment
• Promoted mutual understanding and awareness between the business side and the technical side of the enterprise through web app, dashboard, executive report and data palooza. Tailored the model/solution to truly fulfill the demands of the stakeholders.

• Extracted Medicaid and other public health data from the data warehouse, census and social media for stakeholders to oversee the overall picture about the Medicaid population in Indiana
• Ran descriptive and inferential statistics to help stakeholders gain insight of certain public health trends such as the opioid crisis
• Fitted data into machine learning models (e.g. NLP, ARIMA, random forest, neural network etc.) to predict incoming trends of certain issues (e.g. Medicaid cost, potential Alzheimer's Disease patients etc.) so that necessary legislative actions could be undertaken in advance

Developed a R shiny app to help physicians interactively visualize sample cancellation data (over 270 million observations). Identified sites in need of targeted training and was also used to track their progress towards improvement. Increased site & sponsor interactions and reduced cancellation rates.