Kunjbihari .

Converted unstructured qualitative feedback from PDF documents into clean CSV files using automated extraction with PyMuPDF. Represented feedback text as sentence embeddings, applied UMAP to reduce dimensions to 100, and used K-Means clustering to group similar responses. Tuned clustering parameters and evaluated results using the Silhouette score (0.95) and the Davies–Bouldin index (0.15). Analyzed the sentiment and emotional tone of feedback using transformer-based language models. Assigned faculty members to appropriate roles by comparing the semantic similarity of feedback embeddings and applying the Hungarian algorithm.

Collected and cleaned 1.6M Instagram posts and 35K firm breach records using web scraping and official APIs, ensuring consistent formatting, removal of noise, and alignment across multiple data sources for downstream analysis. Improved company name accuracy and dataset linkage by applying named entity recognition, fuzzy string matching, and manual validation against the U.S. Firm registries and stock ticker lists. Analyzed caption sentiment using VADER and FinBERT, comparing emotional signals with breach events, and firmlevel indicators to study patterns and relationships across datasets.

Exit From IMSc