Ansh Gupta

Developed a lightweight voice activity detection model (SemanticVAD) by fine-tuning Qwen3-0.6B on conversational data for 4 dialogue states (<|start−speaking| >, <|continue−speaking| >, <|start−listening| >, <|continue−listening| >) Achieved 89.1% overall accuracy, with > 90% accuracy on most classes, by leveraging a custom token hidden state pooling method and fine-tuning with a hybrid loss (Cross-Entropy, KL Divergence, Entropy) for robust performance.

Developed a pixel classification model for flood mapping using Sentinel-1 satellite data to track flooded regions Identified crop-damaged pixels in agricultural land clusters by analyzing pre- and post-event Sentinel-2 satellite imagery Conducted time series analysis utilizing NDVI values from remote sensing data to determine optimal harvest date for crops Built and deployed the GIS Surveyor application for claim management using Flask in a production environment

Performed data collection, cleaning and processing, including ETL, aggregation and scraping from multiple databases Trained a GCN (Graph Convolutional Network) model in semi-supervised fashion to predict cancer driver genes in a network Implemented PyTorch Geometric GNN Explainer for model interpretability, providing node and edge importance masks

Developed a generative AI model using fine-tuned GPT-2 to synthesize novel amino acid sequences with enhanced binding affinities for antigens, accelerating drug development and therapeutic advancements Utilized CNN and Random Forest regression models to predict binding affinities and prioritize antibody sequences Shared 100 optimized sequences; wet lab tests showed 80% above benchmark, 10% below but binding, and 10% not binding