Aman Singh Gulati

When fusing heterogeneous modalities for classification, a central challenge is cardinality heterogeneity: modalities often produce token sequences of vastly different lengths, yet standard symmetric fusion wastes attention capacity under this asymmetry.

Graph Neural Networks (GNNs) break down on zero-degree nodes, as message passing requires neighbors. Without interaction history, unseen entities are sub-optimally embedded, leaving them weakly anchored in the latent space, creating a cold-start bottleneck in retrieval. To address this, we propose GRAFT, a factorized architecture that unifies structural and feature transformations into a shared weight space.

Address intelligence in e-commerce demands accurate geocoding and proactive defect detection under strict sub-50 ms latency constraints. These tasks are inherently coupled: precise spatial grounding provides a strong prior for defect propensity, yet prior approaches optimize them independently. While generative LLMs offer rich semantic representations, they lack spatial inductive bias and fail to meet real-time serving requirements.