Mountain View, California, United States
Applied Science in Pinterest's Advanced Technologies Group
Fully automated customer support chat agent for Google Ads Sample efforts: 1. Product Policy Adherence: Ensure Chat Agent correctly handles user queries involving certain Product Policies. - Delivered best performing candidate as evaluated by in-house expert human raters. - Roadmapped and executed experiments (incl. prompting, CoT, RAG, Instruction Tuning, Reinforcement Learning) - Created Instruction Fine-Tuning (IFT) Data Mixture of Product Policy Tasks to support the experiments - Led a team to establish a LLM rater that rates conversations against a given Acceptance Rubric for evaluation. 2. Improve In-Domain Factuality: using Search over internal corpus for RAG - Impact: lifted factuality metric (% claims supported) from 68.7% to 72.2%
I worked on Deep Learning models that rank results in YouTube Search, the world's second largest search engine. Launched many projects to improve search ranking models (Propose, design, implement, test, iterate, launch.) Impact: +0.15% Daily Active Users (DAU), +0.37% Trail Watch Time, +1.00% Search Long CTR, +0.38% Search CTR. - Collaborated cross-functionally with research, infrastructure and product management teams. Sample efforts: 1. Generative AI for search retrieval: For recommendation-seeking queries e.g. 'country music', use LLM to parse & summarize Google Search results into recommendations to insert into YouTube search results. - Impact: +2.40% Official Music Video Views, +0.99% YouTube Music Views, +0.24% Voice Search Watch Time 2. Improve ML Ensemble Learning: Using a gradient-boosting inspired methodology to improve the performance of an ensemble of models where not all will trigger for every query and learning signals tend to be highly correlated. - Impact: +0.07% Daily Active Users (DAU), +0.51% Long CTR 3. Improve Deep Learning Model for Shorts Ranking: Input & target feature changes; add shorts-related data sources - Impact: +0.37% Sessions with Shorts View, +0.75% Shorts Shelf Long CTR
Built a system to recommend images for users seeking certain kinds of visual inspiration. 1. Built general-purpose end-to-end pipeline to crowdsource data, do feature engineering and train and evaluate Decision Tree & Neural Network models. This pipeline: - provides a unified end-to-end pipeline for gathering data from multiple sources, training multiple types of models and offers multiple model deployment options - has since been used by Googlers on other projects (enabled by detailed documentation) - decreases crowdsourcing job generation time from hours to minutes - eliminates the need to implement certain features by providing new pipeline pathways - adds new sources of (model input) data - improves performance and design in many other areas 2. Developed (and wrote documentation for) image recommendation model, using this pipeline, that outperformed Google Image Rank on an 'inspiration-based' recommendation task
- Managed the technological integration of PredictiveHire (an AI acquisition) into Hudson’s traditional recruitment business - Navigated process change across both companies to launch to thousands of users - Conducted Australian ‘temp work’ market sizing analysis (by industry and agency) to quantify benefit from entering a partnership with Expert360 (an online recruitment platform), which led to successfully closing a deal https://www.afr.com/technology/expert360-s-big-career-move-500m-hudson-deal-fuels-global-plans-20191030-p535ul