Research Projects

🖥️ Featured Projects

Proposed an efficient affinity-driven MTL framework that pre-selects task groups to maximize MTL performance gains.
Proposed task-affinity–driven grouping strategies, improving MTL gain prediction accuracy, and outperforming baseline MTL gain prediction methods (TAG, MTGNet, GRADTAE) by a wide margin, achieving up to 4x higher correlation with ground-truth gains across computer vision, tabular, time-series, and transportation benchmarks.
Achieved 15% lower prediction error than single-task models and ~7% MSE improvement over Naive baseline MTL. Validated on real-world public transit ridership data, demonstrating consistent performance gains across task groups.

Designed and developed a framework for predicting energy consumption for various transit vehicle types using multi-month sensor data, outperforming classic learning algorithms (decision trees and linear regression) by ~ 33% in MSE reduction.
Achieved <5% prediction error for 6-hour trips, by aggregating sample-level predictions across time-series segments.
Built a decision tree–based map-matching module linking noisy GPS to road geometry with 90% accuracy, enabling elevation and distance features.
Models trained on multi-month, multi-vehicle telemetry data (6 vehicle types across 8 months); prediction outcomes informed energy-aware routing and scheduling strategies for transit agencies in Tennessee.
Integrated into a distributed ML pipeline for real-time energy prediction

Developed a neural architecture and feature search framework for route-specific ridership prediction using Automatic Passenger Count and weather data - jointly optimizing prediction error and model complexity
Evaluated on real-world transit ridership data across ten routes, showing that route-specific neural network architectures and features outperform generally optimized models in prediction accuracy (9% lower error).
Architectures optimized per route-task consistently yielded the best results compared to hand-designed baselines, demonstrating the benefit of customizing both model complexity and feature set to task-specific patterns in real-world transit systems.

Analyzed 3.3M+ transit boarding events from Nashville and Chattanooga by integrating farebox, GPS, and telemetry data; performed temporal, spatial, and socio-economic analyses to assess ridership declines across demographics, locations, and time-of-day patterns.
Identified persistent COVID-19 impacts on transit accessibility, informing transit agencies’ strategies for equitable service restoration; findings published in Transportation Research Record.
Findings informed adaptive strategies for transit agencies to prioritize equitable service restoration and plan for future disruptions.

Investigated security concerns and awareness in the smart contract developer community by analyzing Q&A discussions, blog posts, and associated source code from multiple platforms.

Examined Chromium bug bounty program data, including activity logs and rules descriptions, to characterize participant incentives, behaviors, and vulnerability reporting processes.

Languages: Python, Java (basic), C++ (basic), SQL, Bash
ML Libraries: PyTorch, TensorFlow, Keras, Scikit-learn, NLTK, SciPy
Data Analysis Libraries: Matplotlib, NLTK, Scipy, Pandas, Numpy,
Development Tools: Git, Docker, Jupyter, VSCode, AWS (basic)
Research Methods: Automated Machine Learning, Multi-Task Learning, Deep Learning, Neural Architecture Search, Exploratory Data Analysis, Statistical Modeling, Energy Optimization, Transit Ridership Prediction, Smart Contract Security