Selected Projects

A compilation of my research and engineering projects in Deep Learning, Computer Vision, Generative AI, and Graph Neural Networks.

Multi-Sensor Cross-View Panoramic Synthesis

Google DeepMind

Advisors: Prof. Vinay Namboodiri, Dr. Vaibhav Rajan

• Built a diffusion model synthesizing physically consistent panoramas from solely low-resolution satellite data.
• Internalized HR structural priors via a cyclic critic, eliminating explicit geometry needs at inference.
• Bridged polar aerial and spherical ground geometries using a coordinate-aware cross-attention mechanism.
• Outperformed explicitly HR-conditioned baselines across all metrics on a global dataset of 550K images.

---

Cross-View Generative Fusion

Google DeepMind

Advisors: Dr. Vaibhav Rajan, Prof. Sara Beery

• Pioneering unified cross-view generative fusion to overcome fine-grained detail loss in decoupled methods.
• Bridged a critical data gap by curating a novel dataset temporally aligning Sentinel-2 with agricultural Street View.
• Architecting discrete diffusion models to resolve fundamental continuous-discrete latent incompatibilities.

---

Single Source Open Domain Generalization

IIT Bombay

Advisors: Prof. Biplab Banerjee, Prof. Gemma Roig

• Developed SODG-Net, an end-to-end network outperforming benchmarks by 1-14% in open-domain settings.
• Proposed a style synthesis block generating diverse statistical features to effectively simulate novel domains.
• Formulated novel weight learning and margin objectives to establish distinct representations for open classes.

---

Crop Classification & Data Curation

Google DeepMind

Advisor: Alok Talekar

• Leveraged Gemini to semantically filter Street View imagery, isolating optimal candidates for expert annotation.
• Established comprehensive annotation guidelines to curate a high-quality, ground-level crop dataset.
• Developed a pipeline augmenting scarce satellite data with robust, timestamped, Street View-derived crop labels.

---

Graph Neural Networks for Free-space Mapping

MBRDI

Advisor: Dr. Laxmikant Sahoo

• Proposed a novel GNN architecture formulating free-space mapping as a scalable edge-prediction problem [Patent].
• Architected a dual-path Graph Attention Network to independently process node and edge features.
• Implemented class-aware edge retention via feature fusion, achieving parity with inflexible rule-based legacy systems.

---

Incremental Domain Adaptation

IIT Bombay

Advisors: Prof. Biplab Banerjee, Prof. Gemma Roig

• Devised a robust solution establishing discriminative, domain-invariant feature spaces under shifting distributions.
• Retained critical prior learnings by aligning model outputs and batch-norm parameters for representative proxy data.
• Leveraged center loss alongside standard cross-entropy to significantly enhance model prediction confidence.