Working on developing state-of-the-art deep learning approaches to entity matching and entity reconciliation for the Yahoo Knowledge Graph.

Designed, developed and deployed low-latency algorithms for extreme multi-label classification (XMLC), enabling large-scale interest-based / conversion-based audience targeting in a real-time setting.

Developed a multi-scale graph embedding approach for extreme multi-label classification (XMLC), aimed at selecting relevant items from a large number of possible outputs, while automatically categorizing the outputs into hierarchically nested groups. Apart from demonstrating superior performance compared to other factorization machine-based models on public benchmark datasets, the approach was also leveraged for joint conversion prediction across hundreds of predictive audiences.

Developed a Time-Aware Sequential Autoencoder (TASA) to generate time-preserving user embeddings from sequences of user activities, irrespective of the downstream task. The generated embeddings were intended to be seamlessly applicable across different user-related prediction tasks, thus saving redundant efforts that some of the company teams put into task-driven feature engineering. TASA has been integrated into a pipeline designed to provide time-aware user embeddings as a service, and the use of its embeddings obtained lifts in conversion prediction performance on four different user audiences.

Worked on two computer vision projects: 1) Design and implementation of a face recognition system based on a multinomial classification ensemble of neural networks; and 2) Real-time 3D object segmentation using Microsoft Kinect.