• Working on pretraining for the Falcon LLMs on multi-node GPU clusters.
• Writing custom Triton kernels for core transformer layers to speed up training.
• Working on Falcon-Audio, an audio-language model for speech, music and sound understanding, published at ASRU 2025.

• Owned real-time ML systems for Swiggy's search, from modelling to production monitoring, measured on NDCG, CTR and search-to-order.
• Shipped a query intent classification and disambiguation model, a pairwise ranking model for dishes and restaurants, and a Siamese retrieval model.
• Recognised with the Most Valuable Player Award.

• Built an NLU query-matching system that maps customer-support questions to FAQs.
• Implemented Non-negative Discriminative Feature Selection, a manifold-learning approach to unsupervised feature selection.
• Fine-tuned BERT and RoBERTa for document and user embeddings in personalised email marketing, at 165 million data points.

Built a pair-analysis micro-service for IBM InfoSphere Master Data Management that samples misclassified record pairs, shipped as Dockerised REST APIs.

Video captioning (ResNet-50, I3D) and weakly supervised step segmentation (NN-Viterbi) on instructional cooking videos, contributing to an IEEE TPAMI 2022 paper. Code →

Multilevel multiple imputation on longitudinal data, preserving within-group correlation for linear mixed-effects models. Code →