Aug 2022 – May 2025 · 2 yrs 9 mos · Menlo Park, California, United States · On-site

• Led evaluations for multi-modal LLMs, conducting 100+ ablations for LLama 4, 3.1, 3.2 and 3.3
• Crafted synthetic data to improve tool calling, web search & coding performance for LLama 3 405B
• Developed a shared data pipeline framework and automated 50+ evaluation tasks across 4 teams for Llama
• Built contextual Ad2Ad ranking model for Instagram Feed/Story and Facebook Feed, increasing Ads revenue by 0.06%
• Improved Ads ranking models for organic contextual ads based on interacted posts, decreasing eval NE by 0.25

May 2021 – Jul 2022 · 1 yr 2 mos · Bellevue, Washington, United States · On-site

• Improved question answering in Bing Search and Edge browser, increasing F1 score by 75% and relevance in human evaluations by 0.7 percentage points.
• Synthesized 500M+ samples across 15+ languages and trained 4k-context Transformer-based teacher model for above

Sep 2020 – Dec 2020 · 3 mos · Seattle, Washington, United States

• Retrieve videos from text queries using multi-level embeddings from video to represent global, object and motion features.
• Introduced a novel triplet loss using continuous labels, increasing Recall@10 on Berkeley Deep Drive-X Dataset.

Jun 2020 – Aug 2020 · 2 mos · San Jose, California, United States

• Deployed a system to detect people who use mobility aids using on-device cameras in autonomous vehicles
• Filed a patent to find optimal workflow for Deep Learning tasks in edge devices, dynamically choosing cloud server, model and size depending on workload; and achieved mAP@0.5 of 84% on Mobility Aids Dataset with >20 fps

Mar 2017 – Aug 2019 · 2 yrs 5 mos · Seoul, South Korea

• Introduced a method for image retrieval by ensembling features from multiple deep learners using attention masks, achieving state-of-the-art results on SOP, CARS and CUB datasets (published in ECCV 2018)
• Designed and deployed image-based product search for Bixby Vision using Caffe and Tensorflow, optimised for low runtime and memory, currently running on 100 million+ smartphones (Galaxy S8+)
• Developed knowledge-based question answering chatbot for Samsung Customer Service using BERT
• Won first prize in Perfect Product Image Recognition Challenge, presented at ACMMM 2018

Sep 2015 – Feb 2017 · 1 yr 5 mos · Seoul, South Korea

• Designed a conditional rule engine and predictive battery statistics module for Tizen Operating System
• Created apps for abstractive email summarization and contact and calendar event extraction from Emails
• Built a Swagger Codegen module for Tizen to create client SDKs from OpenAPI specification

May 2015 – Aug 2015 · 3 mos · New Delhi, Delhi, India

• Research Paper Recommendation
• Designed an algorithm to recommend research papers to read based on learning aim of a reader.
• Built a topic dependency graph of 500+ topics, based on Wikipedia page links, paper references and textbook glossary.
• Used the topic dependencies along with prior knowledge and reading list of the reader for paper reading suggestions.

May 2013 – Jul 2013 · 2 mos · Bengaluru, Karnataka, India

• Patented a method to increase accuracy of digitisation of handwritten document using crowdsourcing, and co-developed an Android app.
• Devised and used metrics for transcriptions similarity, workers’ performance and language models-based likelihood.