San Francisco Bay Area
Machine Learning Engineer at Google who led the launch of the first-ever search results page built entirely with Large Language Models (LLMs). My expertise lies in productionizing advanced AI, from fine-tuning Gemini models with LoRA and building Retrieval-Augmented Generation (RAG) systems to creating novel ranking algorithms that significantly improve user satisfaction. This is complemented by my experience in architecting scalable backend systems, building high-throughput caching pipelines, and developing foundational frameworks to optimize latency and performance for millions of users.
AI Overviews expansion - Led AI Overviews expansion into key Search Verticals (e.g., Sports, Recipes, Finance). - Improved response quality and relevance by developing a scalable framework for vertical-specific prompt updates and Retrieval-Augmented Generation (RAG). AI Organized Search Results Page - Led the team (10 engineers) that launched a first-of-its-kind Search Results Page built entirely with LLMs (https://blog.google/products/search/google-search-lens-october-2024-updates/). - Developed a teacher-student model system, performing Supervised Fine-Tuning on a Gemini−M model and deploying an efficient Gemini−XS + LoRA adapter for production. - Improved overall user satisfaction by designing and implementing a whole-page ranking algorithm using contextual optimization and Generalized Additive Models (GAMs). - Architected and built high-throughput pipelines for Semantic and Direct caching, significantly reducing end-to-end latency. Recipes Wholepage Experience - Led a team of 4 engineers across diverse projects including personalization, experimentation, feature development, rendering, and latency optimization. - Spearheaded the launch of the Recipes wholepage redesign (0.6% en-US mobile), delivering significant user engagement (+0.52%) through strategic feature ablation and result optimization. W2X ItemSelector Framework - Developed the backend components of W2X, a cross-team initiative to develop foundational frameworks for 1P decision-oriented user journeys (multi-carousel/ filters), impacting ~8.3% of Search traffic. - Designed a horizontal experimentation framework that standardized conditions and significantly improved LiveExperiment quality analysis for verticals from O(weeks) to O(days). 2020 US Elections - Launched Referenda KPs resulting in 15.5M impressions. - Developed a new structured Quotes tab in People OSRP leading to 462M impressions.
" Hierarchical Co-Attention for Visual Question Answering" "Show and Tell" : A Neural Image Caption Generator “Conversation with Shakespeare": A Dialogue Model for Style Transfer using Cornell Movie Dialogues Embedding Words into Vector Space (word2vec & Glove) for word prediction in a context in Python Social network analysis of a multi-attributed network using networkx and Python. Hand written digit Classification using TensorFlow. Consensus in a Distributed Data Management system using Multi-Paxos in Python.
Artificial Intelligence : Topics include intelligent agents, search, game playing, logic, knowledge representation and reasoning, and uncertainty. Data Structures and Algorithms.
Semi-Supervised data classification for enhanced model training to bridge the gap between usage and training dictation data. Weighted TF-IDF features with generative Expectation Maximization model is used to achieve classification accuracy of 83%. Data pipelining and web tool to build an automation system to reduce the analysis time of daily Bixby data by 50% using Spring.
Responsible for developing tooling solutions for Requirement Engineering and conducting trainings on JAMA, V-Model and Requirement Engineering for different teams. Developed a Verification Specification tool for Requirement Based Verification using Python and QT library.