Working under the supervision of Prof.Judy Hoffman

Exploring synthetic-to-real generalization of computer vision applications for aerial imagery and analyzing the effects of intra domain shifts

Designed and implemented image super-resolution algorithms on Sentinel-2 satellite imagery using transformer architecture for the downstream task of improving field boundary detection on agricultural land

Achieved a PSNR value of 36 dB and SSIM value of 0.81 on the curated test set

Performed data cleaning and satellite image preprocessing such as tile splitting, merging and georeferencing, generating feasible Sentinel2 composites with minimum cloud cover

Conducted a study on the feasibility of using GANs and Diffusion Models for the specific problem set

Designed the objective criteria and performed suitability analysis to rank parcels within Henry County based on their suitability for freight related construction and transportation by taking into account environmental, socio-economic, logistical and transportation factors

Designed the decision support tool using ArcGIS to quantify the impact of freight development while channeling growth towards social, environmental and economic resilience

Led the development of a custom passive video liveness solution incorporating a vision transformer-based, multi-channel architecture with optimized frame sampling techniques for detecting Face Presentation At- tacks and achieving a consistent score of over 99% on research benchmarks and in production environments..

Led the AI efforts of the Geospatial team, designing a unique change detection and classification system on large-scale satellite/aerial imagery. Implemented a hybrid approach utilizing segmentation models, automatic imagery alignment, vectorization, and shape enhancement techniques, achieving a consistent F1 score of over 90% across various geographies and significantly improving property tax collection efficiency.

Enabled the FY20 GTM (Go to Market) strategy for Microsoft Public Sector Department by conducting a comprehensive analysis of the smart city landscape in India, performing market research on cloud and AI services adoption, and suggesting a growth-maximizing roadmap targeting industry leaders

Facilitated easy adoption of smart city solutions by understanding the niche technical requirements of each smart city element and designing prototypes using Microsoft products and services.

Designed a computationally efficient deep learning-based semantic segmentation algorithm for self- driving cars that achieved 10x faster speed, 5x less memory consumption, and 224x fewer parameters with similar accuracy compared to the state-of-the-art segmentation algorithms by introducing the concept of transposed depthwise separable convolutions and spark module.

Creating the first dataset of its kind for training HW-NAS latency predictors by expanding the NATS-Bench dataset to include hardware-specific latency measurements and descriptors across various GPUs, CPUs, and Edge devices.

Developing a latency predictor that generalizes to unseen hardware and model architectures without sample measurements on the critical inference path

Analyzed the importance of different VQA components and designed an easy to use lightweight framework that is able to achieve results comparable to state of the art on VQA-RAD dataset

The work highlights the importance of encoder selection and offers a practical solution for VQA in the medical domain

As a member of UC, San Diego, Engineers for Exploration collaborated with San Diego Zoo to develop a real-time machine learning system to monitor and analyze behavioral cues in Aye-Aye for early detection of health issues.

Conducted extensive survey and experimentation on the different Hyperparameter Optimization techniques such as Turbo, Hyperopt, PySOT, and Nevergrad as part of the BBO challenge.

Identified the advantage of trust-region based exploration-exploitation provided by Turbo and the superior candidate selection method (TPE) used by Hyperopt to design a hybrid approach.

Trained neural networks that are robust to post-hoc pruning using Targeted Dropout. The idea was to cut off the snow-ball effect of error buildup in the network’s activations by removing the impact of most weights.

Tested the feasibility of this technique by experimenting with FSGM and PGD attacks on CIFAR 10/100 datasets implemented on ResNet models augmented with Targeted Dropout.

Attempt at developing an automated system to monitor shifting shoreline trends in high-impact regions of the world using satellite imagery (Landsat, SPOT) and deep learning based segmentation techniques.

Achieved an accuracy of 97% in delineating shorelines in sandy beaches worldwide through semi-supervised learning techniques on a custom dataset using a 3 class approach (sea, land, and runup) to counter tidal noise.