Developed RESTful APIs and Java packages for Amazon’s global warehouses inventory controlership management, tracking system with Java, Ant,Jackson, Mockito and AWS in collaboration with 2 cross-functional teams
Research Assistant
Social Cognitive AI (SCAI) Lab at Johns Hopkins University
Responsibilities include:
Pending ICRA pubication, AAAI Symposium pubication accepted: UnclearInstruct: An Embodied Assistance Challenge
Proposed a Spoken Instruction Following through Theory of Mind (SIFToM) model to interpreate acoustic wave and human auditory perception to infer robot goals via Bayesian inverse planning algorithm’ on simulated and real-world data advised by Prof. Tianmin Shu at Johns Hopkins University
Research Assistant
Learning to Defer with an Uncertain Rejector via Conformal Prediction
Responsibilities include:
Proposed a uncertainty-based distribution-free post-train component for learning to defer in enhancing the collaborative performance of human and AI team and rendering safer decisions on tasks ranging from object to hate speech detection via uncertainly quantification, advised by Prof. Eric Nalisnick at Johns Hopkins University
Developed surrogate loss functions in Bayesion sub-optimal approaches for learning to defer problem on wide ResNet, human expert simulators and data augmentation with PyTorch
Proposed active learning pipeline with uncertainty quantification methods including batch ensemble, SNGP, MC-Dropout, and BNN for Wide ResNet on CIFAR-10/100 dataset using TensorFlow and Python
Developed and automated experiments on CIFAR10 w/ corruption, human, Hate Speech, and Street View dataset
Surveyed distribution shift on wide ResNet using OpenCV and visualized using matplotlib and seaborn
Automated and distributed experiments over GPUs with Slurm, Shell, and Docker at Linux HPC
Research Assistant
Augmentation for Distribution Drift in Credit Scoring
Responsibilities include:
Proposed data augmentation algorithms with kernel density estimation against distribution drift of credit scoring models, and improve the AUC of ML models from 0.73 to 0.85 with LightGBM and PyTorch under various economic factors
Surveyed credit risk models in gradient boosting, neural network algorithm with Python, PyTorch, NumPy
Analysis and visualize experimental data statistically with Pandas and matplotlib
Created large-scale databases for ∼2 bn financial time-series data points with Spark and SQLAlchemy