Arbol

• Deployed Google Dynamic World model and Google earth model on Sentinel-2 dataset for improved classification of forest, land, and water, as compared to our existing model. Used Python libraries including geemap, rasterio, and google earth engine to facilitate the implementation.
• Partitioned and downloaded the resulted Sentinel tiles in sections of 256x256 pixels. Then, reconstructed the downloaded parts into a singular tile to enable visualization with QGIS effectively.
• Developed a comprehensive modular function in TensorFlow that accepts a Sentinel-1 ID as input and provides a corresponding Dynamic World TIF file with pixel-level alignment to the Sentinel-2 tile. Reduced manual intervention and streamlined the workflow by up to 80%.
• Currently, building GANs for SAR to optical image transformation and SAR image denoising in PyTorch.
Tools: Python, Pytorch

IIIT - Delhi

• Collaborated with a team of 5 members to conduct a detailed research analysis on types of privacy attacks on machine learning and deep learning models (model extraction, model inversion, membership and property inference).
• Reviewed literature on various defensive measures to protect the privacy and confidentiality of models against different attacks.
• Reproduced black-box model attacks on ML Systems using techniques like ActiveTheif and Knockoff Nets in PyTorch.
• Proposed a technique to build a shadow model that helps in black-box model explanations to explore and understand the behavior of any black-box model in different feature spaces using PyTorch
• Tools: Python, Pytorch

University of Fortaleza, Brazil

• Collected and pre-processed the chest x-ray data for training and evaluation pipeline. Applied techniques such as resizing, normalization, and image hashing algorithm for removing the duplicates.
• Constructed a CNN architecture by adding custom layers to VGG16 model. Fined tuned the CNN architecture for the new Covid dataset.
• Designed and executed an Auxiliary Classifier Generative Adversarial Network (AC-GAN) based model called CovidGAN that generated synthetic chest x-ray images using Keras and TensorFlow.
• Result: Visualized the results with Principal Component Analysis (PCA) and confusion matrix. The addition of synthetic images produced by CovidGAN increased the accuracy of CNN for Covid-19 detection from 85% (F1-score 0.83) to 95% (F1 score 0.95).
• Paper published in IEEE Access.
• Tools: Python, Keras, Tensorflow

Edge Hill University, England

• Trained and evaluated 4 established CNN architectures for corn leaf disease classification: VGGNet, XceptionNet, EfficientNet, NASNet.
• Performed Model Compression and developed an optimized DenseNet model for corn leaf disease identification with Keras and TensorFlow.
• Proved that its performance was close to the established CNN architectures with significantly fewer parameters and computation time.
• Used Grid Search method to find the optimal hyperparameter values and analyzed the models’ performance through rigorous simulations.
• Result: The proposed DenseNet was computationally cost-effective with 98.06% accuracy, 0.07 million parameters and it took 3 minutes per epoch.
• Paper published in Computers and Electronics in Agriculture, Elsevier.
• Tools: Python, Keras, Tensorflow

Languages and Databases

Python

HTML5

CSS3

MySQL

PostgreSQL

Shell Scripting

Libraries

NumPy

Pandas

OpenCV

scikit-learn

matplotlib

Frameworks

Bootstrap

Keras

TensorFlow

PyTorch

Other

Git

AWS

Heroku