Dhruv Shah

• Working on data-driven, personalized recommendation algorithms, as well as item page refinement, to improve the customer search experience.
• Badging on Recommended Candidates: Driving personalized and customized label assignments (e.g. Bestseller, Social Proof, Suited for You) to candidate items for the search and browse pages, with limited real estate, in order to increase customer engagement across various platforms (Web, Mobile Web, and Apps). The goal is to provide item candidates for the search and browse pages, and to design an algorithm to generate the most optimized and diverse label assignments according to customer queries and preferences. This leads to a 1% improvement in GMV, and a 1.5% improvement in ATC.
• Visual Attribute Generation from Item Images: Part-level grounded visual attribute generation from item images for downstream use cases, such as item image cropping for the homepage, and similarity search. For example, generating details such as collar style, sleeve style, pocket style, and print details from product images.

Steered data-driven content intelligence and recommendations for a food app, refining customer experience and optimizing merchant impressions. Led end-to-end initiatives, from data-driven ideation to deployment, fostering continual improvement.

• Home Page Recommendations for Food App: Implemented a two-tower model recommendation system on the app homepage, achieving an 80% VTB surge and a 3% S2B increase.
• Item catalog tagging and taxonomy building: Tagged a catalog of 90 million menu items using LSH
  (Locality Sensitive Hashing) to group dishes, where the top 100,000 root dishes captured 98% of the total
  menu items.
• Merchant Embeddings: Developed item and merchant embeddings using Word2Vec and content-based
  recommendation techniques, enabling diverse recommendation scenarios, such as similar dishes, serviceability and exploring remote restaurants.
• Ads Diversity Solver: Constructed an advanced solution for ad keyword ranking, leveraging the Upper
  Confidence Bound methodology.

Led the development of advanced solutions for customer and merchant servicing teams. Oversaw projects
from classical ML benchmarks to implementing cutting-edge Seq2Seq for key result enhancements.

• Text Summarization for Servicing Mailboxes: Developed a summarizer and label classifier for various
  customer and merchant mailboxes using BERT and T5, which resulted in a 20% average time reduction for
  ticket handling.
• Internal Search Engine: Implemented a Learning to Rank solution for an internal search engine used by
  customer service professionals, integrating RANKLIB with Elastic Search. This initiative decreased the average
  click rank from 3.7 to 2.1.
• Document highlights based on the query: Pioneered the development of a system that generates document
  highlights based on keywords or queries entered by servicing colleagues, significantly expediting the servicing process.

• Employed PDF computation techniques to predict real-time candidate click or booking probabilities for ad selection, contributing to increased Click-Through Rates (CTR).
• Engineered ETL pipelines to compute tailored, aggregated filters for campaign optimization.
• Spearheaded the design, development, and maintenance of large-scale applications, such as Audience and App handling, facilitating real-time bidder functionality.
• Data Transformation Engine: Distributed Transformation Engine using Apache Spark to process streaming transaction data (up to 100 MBPS) to store in a usable format.
• Rule Engine Integration to ETL Pipelines: Integrated Drool Rule Engine into the ETL pipeline, enabling near real-time reflection of business logic changes. Achieved a 50% reduction in ETA for one specific use case.
• Apache Drill Storage Plugin: Implemented storage system that provides real-time stream ingestion and extraction, as well as supports join, group by, and aggregate queries using Apache Drill and Druid.

• Data Transformation Engine: Distributed Transformation Engine using Apache Spark to process streaming
  transaction data (up to 100 MBPS) to store in a usable format.
• Rule Engine Integration to ETL Pipelines: Integrated Drool Rule Engine into the ETL pipeline, enabling near
  real-time reflection of business logic changes. Achieved a 50% reduction in ETA for one specific use case.
• Apache Drill Storage Plugin: Implemented storage system that provides real-time stream ingestion and
  extraction, as well as supports join, group by, and aggregate queries using Apache Drill and Druid.