Projects
Selected research projects across biomedical AI, climate risk forecasting, deep learning optimization, computer vision, and intelligent systems.
AI-based Histopathology 10/2024 – 10/2026
Objectives
Develop end-to-end WSI and multimodal analysis workflows for histopathology.
Responsibilities 9
- Building end-to-end pipelines for whole-slide image (WSI) and tile-level tissue classification and segmentation.
- Developing a RAG-based pipeline integrating PathologyBERT, ChromaDB, Grounding DINO, and SAM2 for text-guided SRC localization and segmentation-based pseudo-annotation of 512 × 512 histopathology tiles.
- Integrating RAG, PathologyBERT, ChromaDB, and CONCH for morphology-guided retrieval and diverse sample selection across nine colorectal tissue classes using 256 × 256 tiles, with QuPath-based verification.
- Implementing autoencoder-based, multi-task preprocessing for pathology images covering artifact filtering, tile extraction, and quality control, to deliver clean, consistent inputs at scale.
- Managing multiple GPU resources to run large-scale image analytics.
- Training and deploying multimodal pathology foundation models (image + text) for analysis and report generation.
- Annotating tumor regions in WSIs with QuPath to produce high-quality ground truth for model development.
- Curating a large tile-level tissue dataset to support benchmarking and reproducible research.
- Developing a patch- and cell-level segmentation model to detect signet ring cells.
AI-Driven Climate Risk Assessment at the Hanford Site 08/2023 – 09/2024
Objectives
Apply AI-based climate forecasting to assess environmental risk factors at the Hanford Site.
Responsibilities 4
- Developing multivariate sequence models (LSTM, GRU, attention) to forecast groundwater contamination.
- Generating realistic sequential synthetic data using GANs and diffusion models to augment limited observations.
- Imputing missing values and subsequences in time-series data to improve model fidelity.
- Implementing Transformer architectures for long-horizon forecasting of environmental time-series datasets.
Optimizer Enhancements for Deep Neural Networks 05/2022 – 05/2023
Objectives
- Improving model reliability by accelerating convergence of gradient-based optimizers (e.g., SGD, Adam).
- Enhancing nonlinear representational capacity using modified activation functions to capture complex relationships.
Responsibilities 5
- Designing and implementing deep learning models to evaluate the proposed methods.
- Integrating the proposed methods into TensorFlow by overriding core modules.
- Developing mathematical proofs and theoretical analyses of the proposed methods.
- Provisioning a distributed compute cluster and a shareable GPU server to support experiments.
- Running experiments and conducting empirical analyses and ablation studies.
Automated Fabric Defect Detection and Visualization 09/2022 – 05/2023
Objectives
Detect and localize abnormalities in fabric products.
Responsibilities 5
- Applied targeted data augmentation to expand the training set and improve model generalization.
- Built a preprocessing pipeline (denoising, contrast enhancement) to improve image quality.
- Developed and evaluated autoencoder-based methods for anomaly localization.
- Benchmarked models on anomaly-detection datasets and performed error analysis and ablations.
- Documented experiments, configurations, and results to ensure reproducibility.
Real-time Fault Detection in Fused Deposition Modelling 07/2022 – 10/2022
Objectives
Real-time error detection in fused deposition modelling using CNNs.
Responsibilities 4
- Collecting sample images across defect categories—staircase, overfill, and void—during fused deposition modelling (FDM).
- Generating additional images via targeted data augmentation to mitigate overfitting.
- Designing a convolutional neural network (CNN) to train and classify defect types.
- Benchmarking the proposed approach against standardized classification algorithms.
Gesture Recognition System for Sign Language and Music Generation 02/2022 – 05/2023
Objectives
- Real-time music generation driven by hand-gesture recognition.
- Sign-language recognition from hand gestures.
Responsibilities 5
- Modified a GRU model to enhance gesture recognition accuracy and performance.
- Developed a validation framework to ensure model effectiveness across various gesture types for maintaining project quality standards.
- Coordinated the technical integration of deep learning models with application interfaces, streamlining real-time translation of recognized gestures into output.
- Managed technical documentation and project guidelines, supporting project scalability and knowledge transfer.
- Facilitated collaboration between technical and user experience teams to optimize interface design for gesture-based applications.
Korean Handwritten Text Recognition 12/2021 – 02/2022
Objectives
Generate Korean text from old-fashioned Korean handwriting.
Responsibilities 4
- Annotated bounding boxes for handwritten text datasets.
- Engineered a preprocessing pipeline using histogram of oriented gradients (HOG) and contrast normalization to extract robust stroke features.
- Contributed to a Bi-LSTM-based OCR model for handwriting transcription.
- Evaluated performance with character error rate (CER) and conducted error analyses to guide model improvements.
Object Detection in Manufacturing and Smart City 03/2021 – 02/2022
Objectives
- Automated detection of illegal parking in accessible (disabled) spaces.
- Automated detection of improperly placed objects (e.g., stickers, badges).
Responsibilities 4
- Collected on-site imagery of accessible parking spaces to build a diverse training set.
- Preprocessed and annotated images to label illegal-parking cases and object placement issues (e.g., stickers, badges).
- Identified and benchmarked suitable deep learning architectures for automated detection, selecting the best-performing model.
- Conducted iterative testing, error analysis, and model tuning to improve detection accuracy.