● Research Innovation: Pioneered Smoothing ADMM approach for non-convex, non-smooth optimization problems with provable convergence guarantees, published in IEEE Transactions on Signal Processing over Networks and IEEE Open Journal of Signal Processing.
● Algorithm Development: Designed robust algorithms for federated learning, quantile regression with non-convex penalties, dynamic graph learning, phase retrieval, and source localization.
● Extended Applications: Developed federated and decentralized variants supporting asynchronous updates, hierarchical learning, and multitask learning scenarios.
● Leadership: Led team of Teaching Assistants for Digital Signal Processing course (Fall 2022), managing curriculum delivery and lab sessions for 100+ students.
Key Projects:
● Smoothing ADMM Framework (PhD Research): Developed novel optimization algorithms for non-convex, non-smooth problems with applications to robust regression, graph learning, phase retrieval, and localization. Extended to federated and decentralized settings with theoretical convergence guarantees.
● Distributed Acoustic Sensing for CO2 Monitoring: Built comprehensive Python library for processing DAS data with applications to Carbon Capture and Storage (CCS). Features include ADMM-based Total Variation denoising with O(N) Thomas algorithm, federated learning architecture for multi-site monitoring, STA/LTA event detection, and F-K spectrum analysis. Includes 44-page technical report with real Ridgecrest M7.1 earthquake data. [GitHub]
● Ultrasound Imaging Toolkit: Developed medical imaging toolkit for breast ultrasound analysis featuring ADMM-based TV denoising, speckle reduction (Lee, Kuan, Frost filters), U-Net/Attention U-Net segmentation, and ResNet transfer learning for benign/malignant classification. Supports DICOM/NIfTI formats and includes convergence analysis with primal/dual residual tracking. [GitHub]
● ABAX Driver Behavior Classification: End-to-end ML pipeline for telematics applications classifying driving behavior (Normal/Drowsy/Aggressive) from raw GPS/accelerometer data. Achieved 100% accuracy with Gradient Boosting using 24 engineered features. Implemented 18 classification models including MCP, SCAD penalties, and MLP neural networks with driver-level cross-validation. [GitHub]
● Secure Large File Upload System: Enterprise-grade Spring Boot 3.2 application supporting 10GB file uploads with Zero-Knowledge End-to-End Encryption. Features AES-256 client-side encryption with PBKDF2 key derivation, streaming uploads, role-based access control, admin dashboard, and full Docker Compose deployment with Nginx reverse proxy and MySQL. [GitHub]
● BibTeX Reference Manager: Java Swing application for deduplicating and verifying academic references. Features PDF reference extraction, smart deduplication by normalized title, online verification via CrossRef/Semantic Scholar/OpenAlex APIs, safe and aggressive correction modes, and undo/redo support. [GitHub]
● Live Football Scoreboard: Java library with Swing UI for managing World Cup scoreboard. Implements Factory and Singleton design patterns, supports real-time score updates, game ranking by total goals, and comprehensive JUnit test suite. [GitHub]
● Federated Learning for Localization: Designed federated smoothing ADMM algorithms for distributed source localization in networked systems with non-convex penalties, enabling robust performance in presence of outliers.
● Dynamic Graph Topology Learning: Developed algorithms for learning time-varying graph structures from streaming data using non-convex penalties, with applications in network inference and signal processing.
● ROS Formal Verification (B.Sc. Thesis): Proposed formal model for deadlock prevention in Robot Operating System message passing, improving system reliability in distributed robotics.
