top of page
No Collections Here
Sort your projects into collections. Click on "Manage Collections" to get started
Projects
Welcome to my projects. Here you’ll find a selection of my work. Explore my projects to learn more about what I do.
TCAP Underground Station Project
As part of my internship with Metro, I collaborated with and led a team of rising high school seniors and college freshmen to create a proposal for the new Atlantic/Pomona Metro station. We took it from initial research to concept development, doing technical drafting and complete Revit modeling. We selected an underground station and developed detailed designs for the plaza, concourse, and platform levels. We incorporated significant safety improvements, like AI cameras, weapon detection, open-door elevators, and clear wayfinding, while also adding community-centered artwork that reflects the area’s cultural history. For this project, I was in charge of the concourse level, leading the group in modeling and creating renders for that floor.
Graph Neural Networks for Real-Time Rail Disruption Prediction and Passenger-Centric Re-Routing
Rail networks frequently experience disruptions that can lead to widespread delays and missed connections. This review examines how graph neural networks have emerged as practical tools for real-time disruption prediction and management in railway operations from 2020 to the present. Unlike traditional forecasting methods that treat stations in isolation, graph-based models can capture the network’s interconnected nature by learning which upstream delays matter and when downstream impacts will arrive. Early deployments demonstrate the tangible benefits: British rail systems achieved improved short-horizon predictions for connections, and Dutch railways discovered that headway dominates delay cascades. France’s national network now runs minute-by-minute forecasts for thousands of trains. The most successful implementations combine several key elements: joint modeling of spatial network structures and temporal patterns, adaptive connectivity, and integration of railway knowledge, such as headways and buffer times, to produce interpretable outcomes that explain why specific connections are at risk. Operational deployment requires matching prediction horizons to decision windows, mixed with robust handling of degraded data quality and careful monitoring for bias against low-traffic routes. The challenges remain, particularly the lack of shared disruption datasets and the limited number of field trials. However, the convergence of improved data feeds, faster inference techniques, and mature graph architecture can make real-time network-aware disruption management a reality. Proper implementation, focusing on operator-relevant metrics rather than abstract accuracy advances, can deliver measurable reductions in passenger minutes lost and missed connections.
IMU Head Sensor
I built a motion-controlled mouse interface by mounting an IMU sensor onto a breadboard and connecting it to an Arduino-equivalent microcontroller, along with two potentiometers for signal filtering. Using MATLAB, I processed and filtered the sensor data in real time to translate head or hand movements into precise mouse-cursor input on the computer screen.
More intelligent access: AI-Driven doors and ramps for inclusive transit
Public transportation is undergoing a massive expansion across the United States, fueled by the recognition of an unsustainable car culture, urbanization, and sustainability concerns. Despite these strides, accessibility features remain outdated, with people using wheelchairs, walkers, and other mobility aids still relying on decades-old systems of slow, manual ramps and uniformly timed doors. This project aims to modernize accessibility features within public transit vehicles by leveraging computer vision and a YOLO-8n-based convolutional neural network (CNN) to analyze CCTV footage and detect mobility aids in real-time. The study benchmarks this approach against a VGG-16-based classifier in single-label scenarios to demonstrate YOLO's robustness for multi-object detection tasks. The YOLOv8n model achieved superior performance with a precision of 0.9946, a recall of 0.9846, and an F1 score of 0.9893, outperforming the VGG16 baseline across all metrics. Upon identifying a mobility aid, the system would signal transit vehicle doors and ramps to remain open or deploy automatically without delay, reducing human error and safety risks while improving accessibility for nearly 70 million Americans living with disabilities.
The Bike Share System of Los Angeles: Is it Ready for the 2028 Olympics?
Los Angeles, over the past couple decades, has been trying to remediate the effects of mid to late 1900s urban planning that created the car-centric behemoth that peaked in the 00's and 10's (Novak, 2013). The main transit agency, Metro, recently received 900 million dollars to improve their transportation network which includes but is not limited to; bus, bikes, light rail, and heavy rail (LACity.gov, 2024). This is important because efficient public transportation networks are crucial to urban sustainability and is a solution to polluting cars (APTA, 2018).
Measuring transit the success of these new initiatives is difficult because of the multitude of factors that go into making a transportation network efficient. Therefore, I will use the 2028 Olympic games as a stasis point to see if LA's transportation network will be ready in time for the summer festival by analyzing the efficacy of bike share programs in LA.
Measuring transit the success of these new initiatives is difficult because of the multitude of factors that go into making a transportation network efficient. Therefore, I will use the 2028 Olympic games as a stasis point to see if LA's transportation network will be ready in time for the summer festival by analyzing the efficacy of bike share programs in LA.
VGG16 & YOLO Object Detection for Cars
In “Object Detection Utilizing AI,” I introduced object detection as the combined task of localizing and classifying multiple objects in images and video, highlighting applications in autonomous vehicles, surveillance, and medical imaging, as well as key challenges in accuracy and efficiency. I explained core techniques such as sliding windows and image cropping and detailed a project pipeline that classified image patches as car, truck, or other, progressing from a simple perceptron (~72% accuracy) to a convolutional neural network (~82%) and finally to transfer learning with VGG16 (~95%). I also discussed YOLO as a fast single-stage detection alternative and concluded by emphasizing object detection’s growing impact on automotive safety and intelligent transportation systems.
bottom of page