Pengcheng Ding

Production Pipeline and Training Framework
Created an end-to-end batch prediction pipeline in Airflow for predicting product tags.
Improved resource utilization on shared Spark clusters.
Implemented automated failure recovery, reducing recovery time from hours to minutes.
Built the pipeline to run on GCP infrastructure, ensuring scalability.
Engineered a scalable training framework for multi-objective image attribute extraction and classification.
Reduced model development time significantly by leveraging raw images or embeddings.
Developed a modular TensorFlow training suite with custom training loops, customizable model classes, and specialized loss functions.
Encapsulated the system within a Dockerized Kedro environment for simplified deployment and enhanced portability.
R&D for Model Advancement to Improve Catalog Data Quality and Product Findability
Developed novel training methods to handle scarce true labels by integrating noisy label pre-training, programmatically generated labels, and human annotations.
Achieved a 20% average precision (AP) increase over existing models (based on CLIP) and a 40% F1 increase against training on noisy labels alone.
Created custom training loops to obtain optimal task grouping in multi-task models.
Led to a 10 percentage point AP increase.
Improved style models to include KNN-based tabular features, increasing accuracy by 2-5 percentage points and reducing the need for frequent retraining.
Optimized image shot angle prediction models by designing label encodings to stakeholder needs, enhancing model reliability for downstream tasks and reducing the severity of prediction errors by 20%.
Collaborated with stakeholders to design A/B tests and verified positive revenue impact.

Developed novel training methods to overcome challenges with scarce true labels by integrating noisy label pre-training, programmatically generated labels, and human annotations.
Achieved 20% average precision increase over existing models based on CLIP and 40% F1 increase against training on noisy labels alone.
Created custom training loops to obtain optimal task grouping in multi-task models by identifying how tasks are impacted by updates to shared model layers.
Led to 10 percentage point AP increase.
Improved style models to include KNN-based tabular features.
Increased accuracy by 2-5 percentage points and reduced the need for frequent retraining due to improved model robustness.
Collaborated with stakeholders to design A/B tests and verified positive revenue impact.
Optimized image shot angle prediction models by designing label encodings to stakeholder needs.
Enhanced model reliability for downstream tasks and reduced the severity of prediction errors by 20%.
Crafted end-to-end batch prediction pipeline in Airflow for predicting product tags.
Improved resource utilization on shared Spark clusters and implemented automated failure recovery, reducing recovery time from hours to minutes.
Built the pipeline to run on GCP infrastructure, ensuring scalability.
Engineered scalable training framework for multi-objective image attribute extraction and classification.
Leveraged either raw images or embeddings.
Reduced model development time from days to hours for processing tens of millions of images.
Created a PySpark-based preprocessing pipeline for streamlined image and label processing into tf-records format.
Developed a modular TensorFlow training suite with custom training loops, customizable model classes, and specialized loss functions utilizing both CNN and Transformer backbones.
Encapsulated the framework within a Dockerized Kedro environment for simplified deployment and enhanced portability.

Designed pricing incentives to affect supply-demand equilibrium of power markets to recover infrastructure costs.
Simulated electricity transportation network and its effects on market equilibrium.
Determined optimal network expansion decisions to correct distortions from these prices in game theoretic bi-level optimization models.
Built solution heuristics and solved models using HPC clusters.
Aided World Bank to finance its multi-nation infrastructure investment by allocating benefits of international trading and planning using cooperative game theory in optimization models.
Calculated Shapley values.

Assisted Alphabet Health to increase client accessibility to Multiple Sclerosis literature by building their first keyphrases extraction models for literature abstracts.
Implemented LSTM and Bi-LSTM-CRF sequence-to-sequence named entity recognition models for extraction.
Utilized multiple embeddings to improve performance.
Built Elasticsearch backend for storage and clio-lite based Streamlit front-end to conduct contextual search.
Enabled training of extraction models on any search term on PubMed.
Augmented abstracts with extracted key phrases through end-to-end pipeline implemented in MLflow.