Implemented hierarchical 3D Swin Transformer with shifted window attention for multi-modal brain MRI segmentation on BraTS-2020, achieving 0.905 Dice on whole tumor detection through four-stage feature extraction (depths [2,2,2,2], heads [3,6,12,24]), composite Dice-CE loss optimization, and integrated Grad-CAM visualization revealing progressive abstraction from local texture to global semantic tumor representations across 369 training cases.

Systematic investigation of double descent phenomena in social media domain adaptation across GPT-2/LLaMA/BERT families (124M-65B parameters), examining performance degradation at medium scales and recovery at large scales through five tasks (sentiment analysis, hate speech detection, stance detection, sarcasm detection, topic classification) on multilingual Indian social media datasets, deriving PAC-Bayes generalization bounds and platform-specific scaling laws via Neural Tangent Kernel analysis and information-theoretic noise characterization.

AI-powered bioinformatics pipeline integrating DistilGPT2, GPT-Neo-125M, and DialoGPT for automated sequence optimization across 20 specialized tools including DIAMOND alignment, InterProScan domain analysis, ViennaRNA structure prediction, oxDNA molecular dynamics (1000ns trajectories), COPASI metabolic modeling, IEDB epitope prediction, and DNA Chisel constraint optimization, processing SARS-CoV-2 spike mRNA through systematic genomic annotation, codon usage refinement, and immunogenicity assessment with comprehensive quality scoring across GC content, structural complexity, and translation efficiency metrics.

Implemented a novel framework unifying Neural ODEs with adversarial flow matching and Bayesian inference, introducing trajectory-level discriminative training, low-rank tensor parameterizations for computational efficiency, and variational posterior distributions for calibrated uncertainty quantification in continuous-depth dynamical systems.