This project focuses on automated traffic sign recognition using a deep learning model trained on the GTSRB dataset. A Convolutional Neural Network (CNN) is implemented in PyTorch and deployed via a Streamlit web app, allowing users to upload and classify multiple traffic sign images at once. The system provides real-time predictions with confidence scores and displays results in a grid format for easy interpretation.
Bicycle Level of Traffic Stress (LTS) is a data-driven framework that objectively measures how comfortable a road feels for cyclists, with levels ranging from LTS 1 (low-stress, suitable for most cyclists) to LTS 4 (high-stress, requiring high confidence). Crossing Level of Traffic Stress (xLTS) evaluates the comfort and safety for cyclists at intersections or crossings, considering factors like volume and speed of the traffic being crossed, crossing distance, presence of a bike signal, crossing island, RRFB, etc., to rate the stress level Many cyclists feel anxious about crossing busy intersections, regardless of how comfortable the road segment is. By including xLTS, we can better understand the overall comfort level for cyclists, leading to more informed infrastructure decisions.
The introduction of electric bikes to Boston’s BlueBikes system has likely altered trip dynamics, including trip durations, demand patterns, and station usage. This analysis explores the impact of e-bikes on these factors, focusing on differences in trip lengths, popular routes, station turnover, and mode choice between electric and traditional bikes. By comparing e-bike trips with traditional bike trips, the goal is to understand how e-bikes influence overall system performance, user behavior, and mode preferences. These insights will help inform strategies for optimizing bike distribution and enhancing the efficiency of the bike-sharing system.
I created a multi-user web app that lets each user log in with a unique username and record their sleep sessions. All entries—start and end times, duration, and comments—are automatically saved in the browser’s Local Storage under that username. The app also generates a daily total sleep chart using Chart.js, providing an at-a-glance visualization of sleep patterns. Users can edit or delete existing records and export or import their data as JSON files, making it easy to back up or transfer. This simple, self-contained tool helps maintain ongoing sleep logs without needing a server or database.
The Four-Room Reinforcement Learning (RL) problem is a classic grid-world environment where an agent navigates through four interconnected rooms, separated by narrow doorways. The agent must learn optimal policies to reach a goal state from any starting position, using actions like moving up, down, left, or right. The problem is designed to challenge exploration and learning efficiency, as the agent encounters varying rewards and obstacles, requiring it to balance exploration and exploitation while navigating through a structured, partitioned space.
On Route 39, the MBTA’s fourth-busiest bus route, buses experience an average delay of 43 s at the Perkins and Bynner Street intersections along South Huntington Ave. While simply implementing Green Extension, a Transit Signal Priority (TSP) tactic, alone provides limited bus delay reduction, combining it with a TSP-friendly signal control logic and a TSP-friendly intersection layout cuts the delay down to just 5 s. This strategy also reduces pedestrian wait times by more than 60%, without sacrificing bike lanes or parking. Microscopic traffic simulations using PTV Vissim reveal that this smart approach does not even increase auto delay.
Floating bus stops, adjacent to bike lanes, require bus passengers to cross the bike lane to board or alight, creating challenges, especially for visually impaired and wheelchair users. This study examined different floating bus stop designs in Greater Boston, analyzing rider behavior and conflicts with bicyclists, and provides design recommendations to improve accessibility and safety.
Published in 3rd IRF Africa Regional Congress & Exhibition, 2023
Raveling, the loss of aggregate from surface dressing, poses significant maintenance challenges and economic burdens for road agencies, especially in developing countries. This study explored the use of locally available waste materials, such as ground-tire rubber (GTR), as sustainable bitumen modifiers to mitigate raveling, finding that GTR-modified bitumen significantly reduces aggregate loss under adverse conditions, making it a viable alternative for improving chip seal performance.
Published in Office of Transportation Planning, Massachusetts Department of Transportation, 2024
Floating bus stops, adjacent to bike lanes, require bus passengers to cross the bike lane to board or alight, creating challenges, especially for visually impaired and wheelchair users. This study examined different floating bus stop designs in Greater Boston, analyzing rider behavior and conflicts with bicyclists, and provides design recommendations to improve accessibility and safety.
Published in GitHub Repository, 2024
I conducted a detailed analysis of Bluebikes (the public bike share system in the Metro Boston area) data to understand how the introduction of electric bicycles influences transportation choices and urban dynamics. By applying various statistical models, I assessed the shifts in demand for public bike-sharing systems, examining factors such as user preferences and usage patterns.
Published in GitHub Repository, 2024
Developed an integrated framework combining predictive modeling and optimization techniques to enhance the efficiency of electric vehicle charging infrastructure. Leveraged the Adaptive Charging Network dataset from Caltech, containing extensive charging session logs and other relevant data.
Published in Expert Systems with Applications, 2025
We propose SAAM-ReflectNet, a deep learning framework that automates traffic sign detection, classification, and retroreflectivity estimation by integrating robust spatial-semantic feature extraction, Sign-Aware Attention, and multimodal fusion of RGB and LiDAR data. Achieving a mean Average Precision (mAP) of 0.635 and RMSEs of 0.169 (foreground) and 0.147 (background), ReflectNet demonstrates scalability and accuracy, making it ideal for large-scale, proactive traffic sign maintenance.
Published in ASCE International Conference on Transportation and Development, 2025
This paper summarizes global best practices for integrating separated bike lanes with bus stop infrastructure, analyzes bicyclist and transit rider behavior at various floating bus stops to understand their interactions, and develops design guidelines to improve the safety and accessibility of floating bus stops.
Download Download Slides
Undergraduate course, Northeastern University, Boston, MA, 2023
Tutored fundamental concepts of Newtonian physics to practical applications in solid mechanics, equilibrium of particles and rigid bodies in various dimensions, internal forces in trusses and frames, creation of shear and moment diagrams for beams, and computation of stresses induced by various forces, moments, and torque.
Study abroad program, Technische Universiteit Delft, The Netherlands, 2024
Assisted in teaching Vision Zero traffic safety principles and design treatments for safe roads, intersections, and crossings; and high-quality transit service planning and design. Other topics explored are traffic-circulation planning and policies to promote safety, prevent roads from becoming barriers to walking, cycling, or transit, and to create car-free and car-lite zones.