Research projects, collaborations, and independent work.
Full-Waveform Inversion for Seismic Velocity Prediction
Ongoing
Research Supervisors: Naveen Gupta & Divakar Vashisth
Extending advisor’s ICLR 2025 latent-space-translation framework for subsurface forward/inverse problem to using Diffusion priors and flow matching
Working on optimal-transport-based misfit functions (Wasserstein distance, unbalanced OT) as an alternative for L2 loss to mitigate cycle-skipping on Marmousi/Overthrust benchmarks
Literature survey on diffusion- and flow-matching-based generative priors as regularizers for Bayesian FWI
CSD722 Project: Depth-Conditioned Video Generation using ControlNet & AnimateDiff
Jan – May 2026
Research Supervisors: Dr. Sumit Shekhar & Dr. Saurabh Shigwan (CSE Dept, SNU)
Extended ControlNet to text-to-video diffusion models using AnimateDiff and Motion LoRA
Modified model architecture to prevent training collapse where generated videos ignored depth-map conditioning; stabilized training by adding auxiliary supervision losses
Conducted ablation experiments on prompts, depth conditioning, and text guidance
Conducted time series pattern recognition and sentiment analysis in US financial markets using 7 years of daily data across 31 financial indicators to model 2, 5, and 10-year bond yields
Applied PCA and Mutual Information (MI), identifying 10 key indicators explaining 95% of variance; optimised cluster timelines using rolling windows to identify dates with maximum regime separation
Best Silhouette Score: 0.45; ANOVA p-value: 1.21 × 10⁻¹⁵⁴, Kruskal-Wallis p-value: 1.25 × 10⁻²¹¹, confirming significant differences across clusters
Research Assistant: Investor Behavior Analysis using Deep Learning
2024–2025
Research Supervisor: Dr. Ashish Vazirani
Developed an unsupervised multimodal model to analyze investor emotions and decision-making patterns in Shark Tank videos
Combined facial expression recognition, RoBERTa-based transcript analysis, and tabular neural networks for financial characteristics, with cross-modal attention to model investor behavior
Fine-tuned EmotioNet for facial expression recognition and RoBERTa for transcript analysis
Generated time-series investment likelihood scores (–1 to 1) for each second, validated through accuracy of final decision