Enhancing user experience in Microsoft Bing by removing Junk (Dead & low-quality) URLs from search results. Fine-tuned and applied knowledge distillation on a BERT-based model, achieving 99% precision and 86% AUPRC for Junk URL detection. Integrated unsupervised clustering with a high-precision Junk classifier, building a scalable platform detecting over 12 billion Junk URLs. Automated URL categorization with a GPT-based LLM, cutting human labeling costs by $100K annually.

Engineered robust Big Data pipelines using Spark streaming APIs and Kafka, enabling low-latency blocking of junk pages. Revamped legacy Junk pipelines using Azure tools (Logic Apps, Functions and ARM), significantly reducing on-call workload. Achieved a 75% reduction in false positives in Junk detection techniques through in-depth analysis of user behavior on Edge browser.

Spearheaded a collaborative effort with fellow-mentors from Google Brain to enhance the open-source TensorFlow Datasets library. Mentored a student to utilize contributions from TensorFlow & Hugging Face communities, effectively doubling the readily accessible datasets.

Pioneered the enhancement of scanned PDF accessibility by adding an Optical Character Recognition (OCR) feature in the Edge PDF reader. Enabled users to select and search text in images of PDF files, enhancing the PDF reader's capability. Created a network communication workflow for making Azure Cognitive Services API calls via the browser network layer.

Developed an open-source HsYAML library for serializing and deserializing YAML documents in Haskell. Extended the data model to allow round-trips while preserving comments, anchors, etc. Achieved 99% accuracy on YAML-Test-Suite, establishing HsYAML as the best YAML processor.

Crafted educational materials to aid students preparing for JEE Main and Advanced exams.