Tianhao Wu

Led and actively contributed as a key member of a core team within the business unit. Collaboratively supported diverse data science teams while spearheading efforts to optimize and enhance the in-house AI platform with VP-level exposure.

• Architected and delivered production-grade machine learning systems on GCP with seamless integration of Airflow, Dataflow, Vertex Endpoint, and BigQuery to support tabular and non-tabular models.
• Developed standardized containers supporting a range of machine learning frameworks (Tensorflow, Pytorch, Transformers, etc.); defined serving interface to streamline model development and deployment workflows.
• Led the design and implementation of continuous training pipelines by leveraging Kubeflow. Collaborated with Data Scientists to establish model monitors, triggers, and alerts to ensure the model performance and accuracy.
• Collaborated in the design and implementation of a streamlined front door layer for the in-house AI platform, automating deployment workflows and resource creation; reduced develop-deploy cycle from days to minutes.
• Participated in corporate LLM roadmap and use case definition; led the design on LLM applications and systems on GCP; contribute to internal AI Agent initiatives using LangChain and Vertex AI.
• Built and maintained model development Docker environments integrated with Jupyter, MLflow, CUDA, GPU devices, etc; defined model development repository template bridging deployment technical gaps.
• Led junior colleagues on PoCs, designing ML workflows, supporting production environments, and conducting code reviews. Mentor interns and coops on scoping and developing short-term projects.
• Managed contractors in projects to keep deliverables inline with design and requirements; conducted code review and led the development and deployment of project components and features.

Worked cross-functionally with Risk and Tech team on risk management and machine learning solutions including running experiments, implementing model training/serving pipelines, and deploying models into production.

• Developed an embedded hybrid machine learning model for pay-date prediction on 80000+ employers; reduced 24% of the defaults caused by inconsistent pay-date and mismatched employers.
• Implemented name normalizing models for employer names from loan applications; increased matching accuracy from 67% to 93% by leveraging NLP, string metrics, and phonetic algorithms (NLTK, Soundex).
• Created customer profiles using graph database (Neo4j) as fraud features; built XGBoost models to detect fraudulent cases; deployed as lightweight PMML models with Java for low-latency predictions.
• Developed various Python-based models and the corresponding ETL pipelines for offline training and online inference; hosted model endpoints as microservices (REST) running in Docker containers.
• Performed data product and feature variable analysis from fraud and default perspective; created automated variable performance reports using Python; transformed the results into business models.

Promoted in 6 months as Technical Project Lead. Worked in a 4-member agile team on end-to-end projects tailored to clients' business requirements. Collaborated with Samsung, fin-techs, and AI start-ups.

• Developed machine learning projects designed for extracting keywords from semi-structured receipt text data; analyzed 32000+ receipts and optimized model performance on 80% of total receipts.
• Increased extraction accuracy from 28% to 82% by applying solutions combined with supervised-learning, NLP and fuzzy-matching algorithms; increased average app engagement time by 30%.
• Improved ML pipeline efficiency by utilizing Big Data applications (Spark, HDFS); reduced model training from 2 hours to 5 minutes, and created automated dashboards to monitor model performance.
• Leveraged AWS cloud technologies (S3, EC2, EMR, Lambda, SageMaker) to create and maintain serverless Python REST APIs for data transformation and model deployment.
• Lifted sales forecasting accuracy by 4-5%; developed automated Python API solutions to integrate models for 13 products; re-designed ETL processes and improved efficiency by 70%.