Research & Projects

What I'm Working On

My work is driven by a simple idea: AI should make roads safer and people more independent. That plays out across a few different threads right now - some in the lab, some in the real world, and one that's already in people's pockets. My research interests are in Edge-AI Deployment, Autonomous Vehicle Safety, Pedestrian Behavior Modeling, Cooperative Driving Automation, Computer Vision for Transportation, Assistive Technology, Intelligent Transportation Systems, Digital Twins & Smart Cities

Edge-AI for Visually Impaired Navigation (ApexAI)

At ApexAI, we built an on-device navigation system for visually impaired users that runs four transformer models (SegFormer, DPT, DETR, and a Vision-Language Model) on a standard iPhone at under 200ms - no cloud, no internet required. The system identifies walkable surfaces, detects hazards, locates objects, and describes entire scenes in natural language, all in real time. It's live on iOS and currently pending a U.S. patent on the underlying architecture. We're currently raising a seed round to scale it further and develop our Phase 2 smart glasses.

Driver & Pedestrian Attention with Meta Reality Labs

I'm leading a partnership with Meta Reality Labs to study how drivers and pedestrians actually pay attention in conflict scenarios - using Meta's AI-equipped glasses to capture gaze and attention data in naturalistic settings. We're applying ensemble learning to analyze how environmental variables (lighting, intersection complexity, traffic density) affect attention allocation. The goal is to produce human-behavior benchmarks that AV developers can use to make their systems safer around real people.

Cooperative Driving Automation (CDA) Platform

We're built a 1/10th-scale fleet of autonomous vehicles for cooperative driving research. These aren't toy cars - they run real V2X communication stacks, cooperative perception algorithms, and platooning protocols, bridging the gap between pure simulation and expensive full-scale testing. The platform lets us experiment with multi-vehicle coordination scenarios that would be dangerous or impractical to run at full scale.

PRISM: AI-Powered Real-Time Mapping

Co-PI on a USDOT ARPA-I project developing dynamic mapping systems for safer AV operation. The system integrates real-time environmental sensing with ML-based hazard prediction to enable adaptive route planning - the kind of infrastructure layer that AVs will need as they move from controlled test zones to messy real-world roads.

AV-Pedestrian Safety Modeling

At UBC, I got deep into the question of how AVs and pedestrians actually behave around each other. I built a transformer-based trajectory prediction model trained on 1,500+ hours of nuPlan driving data that cut AV displacement error to 2.07m, compared to 5.89m for baseline approaches. I also developed a multiagent inverse reinforcement learning (MAAIRL) platform that simulates AV–pedestrian interactions across four cities - Boston, Las Vegas, Pittsburgh, and Singapore - achieving 81–84% accuracy in predicting evasive actions like swerving and braking. The cross-city analysis revealed that the competitive dynamic between AVs and pedestrians varies significantly by urban context, with Las Vegas being the most adversarial and Singapore the most cooperative. Published in ASCE Journal of Transportation Engineering and Canadian Journal of Civil Engineering.

Real-Time Traffic Safety Analysis

At UCF, I led the Smart Intersection team on a multi-university NSF project (CS3 Street Scapes) focused on understanding pedestrian behavior at urban intersections in real time. We built keypoint and dense-pose estimation models for predicting crossing intentions, and developed a parallel dense video captioning framework for end-to-end traffic safety analysis. That work placed 6th in the NVIDIA AI City Challenge (CVPR 2024, Track 2, score: 29.01). Our team also published a low-light image enhancement framework for fisheye-lens cameras at the same CVPR.

Interactive Traffic Prediction Platform

My dissertation was about making traffic prediction practical and accessible. I designed a Graph Convolution–Gated Recurrent Unit (GC-GRU) architecture for traffic forecasting using connected vehicle data, and built GPU-accelerated ETL pipelines (RAPIDS) that achieved a 50× speed-up for processing Missouri's entire statewide connected vehicle dataset. The end product was NOCONS — an interactive, speech-driven web application that lets transportation planners query traffic conditions by voice and see visual predictions in real time. Watch the demo →. During this period I also built a YOLOv5 + DeepSORT pipeline for automated retail checkout that won 4th place at the NVIDIA AI City Challenge 2022 (CVPR).

Urban Mobility & Ride-Hailing

For my master's at TUM, I studied how ride-hailing services like Uber were changing transportation patterns in Munich. Using a stated-preference survey of 500 participants and agent-based simulation in MATSim, I found that larger TNC fleets significantly cut wait times without affecting trip durations, while smaller fleets were actually more efficient during peak demand. Published in Transportation Research Record.

Technical Skills

• Languages: [Python, Java, JS, SQL, C, MATLAB, R]

• Frameworks: [Pytorch, Tensorflow, Keras, Scikit, Django, OpenCV, ReactJS, Node.js]

• Softwares: [VISSIM, Synchro, SimTraffic, AutoCAD, Microstation, Carla]