Some of my recent projects

1. Explored the robustness of language models by testing five adversarial attacks on datasets like MNLI, SNLI, QNLI, MRPC, and a financial dataset.

2. Implemented two defensive strategies against these attacks: traditional adversarial training and a novel method developed by our team.

3. Utilized Python3 and Google Colab Pro with Tesla T4 GPUs for modeling and computational work.

4. Incorporated major libraries such as PyTorch, HuggingFace, and Textattack for model training and attack simulation, supplemented by Optuna and Scikit-learn for optimization and analysis.

5. Assessed the efficacy of defenses using models like Deberta v3 xsmall, distilbert base uncased, and a fine-tuned FINBert.

6. Conducted a detailed analysis of attack-defense combinations, providing insights into improving LMs' security against adversarial threats.

As part of my work at the Health Analytics Centre at Georgia Tech.

1. Collaborated with pediatric cardiologists to curate a list of ICD 9 and 10 Diagnosis and Procedure codes.

2. Used MySQL to extract an appropriate study population from the Medicaid Claims (MAX, TAF)  using the curated ICD/CPT codes.

3. Dealt extensively with data quality issues, performed data cleaning and preprocessing, outlier detection, etc. in MySQL.

4. Used R to build interpretable statistical models on the data.

5. Results published at The Journal of Pediatrics. 

Skills: MySQL, Python, R

Life Expectancy is one of the major indicators of how well-developed a country's welfare state is. This project aims to understand what factors majorly influence the Life Expectancy of a country and to what extent. 2014 data including several economic, social, mortality, and immunization factors for most countries was collected from WHO and World Bank sources. A regression analysis was performed on this data. 

Skills: Regression Modeling, R, Python, MS Excel 

The Minimum Vertex Cover (MVC) problem is a classical NP-complete optimization problem defined as follows: Given an undirected graph 𝐺 = (𝑉 , 𝐸) with a set of vertices 𝑉 and a set of edges 𝐸, a vertex cover is a subset 𝐶 ⊆ 𝑉 such that ∀ (𝑢, 𝑣) ∈ 𝐸 : 𝑢 ∈ 𝐶 V 𝑣 ∈ 𝐶. Find a vertex cover of the smallest possible size. 

In this project, we tackle this problem via 4 approaches and compared the performance of these algorithms empirically.

 The four approaches covered are an exact algorithm using Branch and Bound, an approximation algorithm using Maximum Degree Greedy,  and two local search approaches - Simulated Annealing and FastVC

Skills: Algorithm Design, Python, Git

Used the Plotly module to visualize the spread of the Covid 19 virus globally and compare the spread across countries and visualized the impact of national lockdowns on infection spread and death rate in Italy and Germany.

Skills: Python, Data Visualisation

Automated the process of building an orthomosaic map from image and GPS data by developing a script using the Python API of Agisoft Metashape.  Further obtained step-specific time and resource consumption data to understand the computational requirements of this process.

Skills: Python Scripting, API

Converted micro MRI films of the human wrist joint into 3D models to perform Finite Element Analysis. Parts modeled include the bones, cartilages, and ligaments.

Skills: Medical Image Segmentation and Masking, Ansys, Materialise Mimics

Developed the finite element formulation for a thin plate subjected to axisymmetric loading. Used MATLAB to develop an algorithm to solve this formulation to obtain deflections, slopes, radial, and hoop stresses, to perform a convergence study, and to compare the numerical results with the analytical solution. 

Skills: Mathematical Modeling, Finite Element Method, MATLAB