EXPERIENCES

• Evaluated the performance of ChatGPT-4, Gemini, LLaMA 3, and ChatGPT-3.5 in generating UML class diagrams from informal textual descriptions by developing optimized prompt engineering techniques that improved PlantUML code accuracy by 60% 

• Quantified model performance using introduced metrics for subjectivity of class diagrams across 200 textual descriptions. 

Published 3 research papers in ACL/EMNLP (including Outstanding Paper Award at EMNLP 2023): 

• Deception Detection: Curated 266k+ dataset via GPT-3 prompt engineering; evaluated MPNet/ELECTRA/RoBERTa for mask infilling; developed visualization tools (heatmaps, matrices) Columbia, USA 

• AI Text Detection: Pioneered perplexity/burstiness analysis framework for 1k texts; quantified detection thresholds for 8 LLMs (GPT, OPT, BERT, etc.) 

• Fact Verification: Implemented AllenNLP for Semantic Role Labeling with > 95% accuracy on 390k+ Twitter, FakeNews datapoints 

• Conducted research on IoT enabled diabetes detection system using machine learning models, including EvalML, AutoML, and Artificial Neural Network algorithms. 

• Implemented GridSearchCV, a hyperparameter tuning technique, to optimize the performance of the machine learning models. • Demonstrated the effectiveness of the Random Forest algorithm, achieving the highest accuracy in diabetes detection within the system. 

• Performed exploratory data analysis and applied ML models on CICIDS 2019; achieved 90% attack detection accuracy. 

• Used PCA, correlation plots, and visualizations for interpretability and performance tuning. 

• Performed Text Sentiment Analysis using Natural Language Processing and VADER which gives sentiment scores on some rules for the words used, achieving 95% accuracy in sentiment prediction.

• Utilized five Machine Learning Algorithms to predict the output of sentiment analysis results

• Generated WordCloud visualizations to observe the frequency of different words and created tables of the frequency of the top 5 words.