Dhingra Lab (Natural Language Processing Lab):Fall 2023: Training a Large Language Model (LLM) with visual features to build robustness to text perturbations and perform original text recoveryRudin Lab (Interpretable Machine Learning Lab):Spring 2023: Detecting Epileptic Spikes in EEG Data using Interpretable Deep Learning (ProtoPNet)▪ Harnessed interpretable Deformable Prototypical Parts Network (ProtoPNet) with fine annotations to detect epileptic spikes in EEG data ▪ Tuned network to achieve an AUC of 0.94 with 6% improvement over state-of-the-art SpikeNet and added advantage of interpretability ▪ Refactored and unified codebases for ProtoPNet, IAIA-BL, and Deformable ProtoPNet papers into streamlined, user-friendly codebase▪ Wrote unit tests with synthetic data and CI/CD pipeline to verify end-to-end model procedures; utilized Docker to containerize applicationCarlson Lab:Spring 2023: Finding the Average Treatment Effect of Land Use on Temperature using Causal InferenceFall 2022 - Spring 2023: Predicting the Urban Heat Island Effect with Computer Vision▪ Employed a multi-task optimized Res-Net to pinpoint Heat Islands by forecasting humidity and temperature patterns throughout the day ▪ Pioneered novel multi-step causal inference technique for image matching to isolate the treatment effect of land use on temperature

Fall 2023: Grader for Graduate-level Course: CompSci 671, Theory and Algorithms of Machine LearningSpring 2023: TA for Graduate-level Course: CompSci 527, Computer Vision

Technical Product Manager for following projects:Statpad- Live in-game NBA probability tracker using Machine Learning and Monte Carlo sampling (Strong involvement)GTHC- Full-stack application for optimal group scheduling with Twitter and Dark Sky APIs; used by Duke students for Basketball tenting (Moderate involvement)Launchpad- Web application to make internship recruiting easier (Peripheral involvement)

Project: Glean Data Analytics Invoice Processing with Natural Language Processing

Spring 2021: Georgetown Virtual Reality Doctor Visit Simulation with UnityWinter 2020-Spring 2021: Sparrow Lending Website Development with React, recharts, Firebase

Project: Creating Artificial Worlds with AI to Improve Energy Access Data▪ Generated 10000 image synthetic overhead imagery dataset of energy infrastructure harnessing a GAN for image blending▪ Constructed synthetic image generation pipeline with custom preprocessing and patch augmentations to generate 3000 images per hour▪ Improved object detection model average precision in geographic domain adaptation experiments by combining real and synthetic datasets, achieving a 71% closure of the cross-domain performance gap and 8% increase in mean average precision over baseline dataset ▪ Implemented state-of-the-art CycleGAN and CyCADA domain adaptation methods, demonstrating 12% relative increase in mean average precision over state-of-the-art; integrated mixed batch training within YOLO scripts to standardize stochasticity of experimentsTools Used: PyTorch, OpenCV, Python, PIL, pandas, numpy, matplotlib, GANs for Domain Adaptation (GP-GAN, CycleGAN, CyCADA), Git, Colab

Project: Creating Artificial Worlds with AI to Improve Energy Access DataTools Used: PyTorch, OpenCV, Python, PIL, pandas, numpy, matplotlib, GANs for Domain Adaptation (GP-GAN, CycleGAN, CyCADA), Git, Colab

Artificial Intelligence GroupResearch & Exploratory Development Department▪ Engineered, generalized, and optimized pipeline for collection of satellite imagery and coinciding road traffic data used for collection of 15000 training, validation, and testing image tiles from Sentinel-2 satellite with geopandas, rasterio, StackStac, and Weights & Biases▪ Produced 60% relative increase in alignment of rasterized road traffic data with road network data used in training segmentation network ▪ Deployed PyTorch MA-Net segmentation model with Distributed Data Parallel for pixel-wise traffic regression, yielding a 0.77 R-squared ▪ Enhanced network performance with custom loss functions, post-processing, and integration of CRESIv2 speed limit estimation networkTools Used: PyTorch, SegmentationModelsPyTorch, Weights & Biases, pandas, numpy, geopandas, rasterio, matplotlib, QGIS, Git, Jupyter

▪ Built a virtual reality application from the ground up to display a 360-degree video of fire with coinciding sensor data▪ Constructed live graphical displays of temporal and spatial sensor data; designed heads-up display GUI to aid user understanding▪ Accelerated and automated processes of VR application to enhance UX, speed and adaptability for new input videos/data▪ Awarded by Augmented World Expo's Auggie Awards as Finalist for Best Societal Impact for deployed application and presentationTools used: JavaScript, A-Frame, HTML, CSS