Saksham Kumar Sharma

Investigated whether pluralistic alignment in LLMs arises from memorization by conducting technical research on the IMHI dataset (6,000+ mental health instructions) using HPC and ADA servers for large-scale analytics and model evaluation. Collaborated with IBM Research on developing a system that uses LLMs to solve planning problems by automating the code refinement process through feedback-driven iterations, replacing the earlier human dependent refinement workflow and reducing refinement time by 80%.

• Applied machine learning to COVID-19 imaging data, achieving 98.04% accuracy on 10,000+ images for biological image–based detection. • Proposed a multi-stage ensemble architecture using Fuzzy Ensemble Techniques and rank-based methods, improving detection rates by 4% over prior models. • Expanded Cervical Cancer Detection research using Vision Transformers, Federated Learning, and a novel ensemble method, achieving 97.9% accuracy on 4,000+ images.

• Analyzed hyperspectral satellite imagery with 500+ spectral bands, developing data transformation pipelines that processed over 1 million pixels for environmental monitoring and feature extraction. • Built a web-based visualization framework for satellite image processing, enhancing data accessibility by 60% and accelerating urban planning analyses.

Programmed a scalable data extraction and transformation pipeline using Beautiful Soup, Selenium, and Python for 10+ major FMCG and healthcare websites, processing over 100,000 data points weekly. Incorporated Natural Language Processing (NLP) and data modeling techniques to analyze market dynamics, identify key players, and assess trends across 500+ products.