• Collaborating with the Tulsa forensics team to recover historical information from the 1921 Tulsa Race Massacre using spectral imaging and AI-based analysis.
• Developing multimodal transformer models to associate recovered gravestone text with death certificates and geospatial public records.
• Applying spectral imaging techniques to decipher eroded materials including stone engravings, rusted metals, and deteriorated wood markings.
• Designing graph neural network pipelines to reconstruct social and familial networks from fragmented historical data.

Supervised by Dr. Corey Toler-Franklin

• Developed sarcasm detection models on context-free text utilizing LSTM and Bi-LSTM architectures, benchmarking against LLMs (GPT-4o, GPT-3.5) and logistic regression
• Fine-tuned RoBERTa for sarcasm detection (context-free/with context), achieving F1 score of 0.79 on Reddit data; improved performance with stop-word handling, attention mechanisms, and emoji-to-text preprocessing
• Extended research to multimodal sarcasm detection by combining video, audio, and contextual text with LSTM-based binary classification, achieving an F1 score of 0.808 on Mustard dataset

Supervised by Dr. Raja M

• Performed feature segmentation on blood vessels and exudates in retinal images to identify key indicators for Diabetic Retinopathy (DR) analysis through UNET architecture
• Developed a multiclass classification CNN model on the segmented images to assess the severity of DR on segmented images, enabling more accurate categorization of disease stages

Supervised by Dr. Shazia Hasan

• Discovered and established market connectedness using Vector Auto Regressive (Diebold and Yilmaz) and cross-quantilogram models
• Selected most connected markets as features for time series modeling for markets through Hybrid CNN-LSTM Deep learning models

Supervised by Dr. Aqila

• Utilized Ridge Regression and Long Short-Term Memory (LSTM) models to forecast hourly precipitation and temperature data for the city of Chennai. Feature variables included humidity, cloud cover, atmospheric pressure, and time-based factors (e.g., year, month)
• Developed a functional application that uses an API to fetch 50 days of feature parameters for accurate target variable prediction

Supervised by Dr. Elakkiya R

• Researched on gas sensing technologies utilizing nanomaterials, specifically arsenene and phosphorene, to enhance detection capabilities

Supervised by Dr. K.K. Singh

• Designed and conducted experiments to convert garlic and agricultural waste into amino acids and salts using controlled heat and acidic media

Supervised by Independent Collaboration