Jun 2022 – Present · 3 yrs 9 mos · On-site

• ▪ Researched and developed a model from scratch that generated $5B in revenue by acquiring over 1M new businesses for Amazon Business (AB) from Amazon Retail sites, within 3+ years since the model launch.
• ▪ Attained $2.358B in AB long-term revenue and $670M in in-year AB revenue through acquisition efforts, marking a 177.44% year-over-year increase with 472K businesses acquired in 2023, 21.37% from Amazon Retail sites.
• ▪ Achieved a 71% performance enhancement over current state-of-the-art deep-learning Multi-variate Time-Series Classification (MTSC) models with BCIvNext model, presented internally at the Amazon Machine Learning Conference 2021 with a 30% acceptance rate. Researched and developed a Multi-Modal MTSC model for Amazon's BCIvNext project that consumed glance-view and purchase data to predict the likelihood of customers buying for business on Amazon Retail sites and increased model capacity by 60%-70%.
• ▪ Partnered with the Amazon Ads team to integrate the BCIvNext embeddings in the Amazon Ads Doppelganger project for business use cases, which resulted in a 1300% lift vs. 130% lift of Ads GNN embeddings; the project got Amazon Ads SVP attention.
• ▪ Owner of Value of Advertising Partners for Sponsored Ads (VAP) Causal Inference study. Automated Value of Advertising Partners for Sponsored Ads (VAP) Causal Inference pipeline, within two weeks, that reduced runtime from 12 weeks to 1 week. Executed 6400+ individual casual studies, prepared a final report summarizing findings for L8/L10 leaders, and socialized the report with at least 6 teams across Amazon Ads.
• ▪ Mentored mid-level and senior scientists and developers, performed three L6 promotional evaluations, delivered organization-wide technical presentations, and facilitated scientific advancement learning sessions. In Amazon Ads, part of the science promotion panel in my VP org, mentored scientists to deploy models to production using the simplified experimentation to deployment workflow.

Apr 2020 – May 2022 · 2 yrs 1 mo · On-site

• ▪ Researched the Next Generation Business Customer Identification (BCIvNext) model utilizing multi-variate time series classification to detect business-oriented purchasing patterns among Amazon retail customers that streamlined targeted marketing strategies for Amazon Business conversion.
• ▪ Experimented and surpassed the Multi-Variate Time-Series Classification (MTSC) model accuracy with 10X lower runtime complexity using Target Encoding vs. the state-of-the-art Amazon ASIN embedding developed by the Central Team.
• ▪ Invented a new MLOPS platform and Led the automation of ML pipeline operations, with 100+ steps from preprocessing to post-processing in nine countries by employing AWS services like CodeBuild, Step Functions, Batch, nbdev, and Ploomber. Processed 300 million customer's data (approx. 8TB) twice per week through the MLOPS inference pipeline, using 15 p4d.24xlarge machines, very reliably for 2+ years. Adopted this MLOPs solution to new Amazon Ads team and socializing it across Amazon.
• ▪ Deployed production model across nine countries, using 2% to 70% negative data, resulting in tailored marketing messaging for enhanced prospect acquisition through customized model recommendations. Expedited registration in all nine Amazon Business marketplaces without verification by identifying high-potential prospects through modeling techniques.
• ▪ Contributed to critical improvements in AWS’s GPU monitoring and utilization capabilities as a technical advisor to the Amazon Consumer Finance team. Helped develop a sophisticated GPU metrics system available to AWS customers worldwide, including Amazon. Achieved an impressive 90% fleet adoption across over 10,000 instances, generating $10 million in annual savings, enhancing GPU utilization, and reducing deployment failures by 60% across Amazon.

Jul 2015 – Mar 2020 · 4 yrs 8 mos · On-site

• ▪ Redesigned the end-to-end Amazon Business's monolithic, operation-heavy registration, and verification stack, leveraging AWS Step Functions to implement 50+ fault-tolerant steps, resulting in a reduction of high severity issues from 15+ per week to low tens per year. Assigned component designs to team members, offering insightful feedback on design and code for enhanced readability, performance, scalability, and maintainability.
• ▪ Earned leadership endorsement by crafting technical designs, architectural plans, and phased implementation strategies.
• ▪ Introduced operational efficient Amazon Business Registration 2.0 and Verification 2.0 back-end service stacks worldwide over 2.5 years by re-designing legacy Registration 1.0 and Verification 1.0 monolithic systems, processing 25K+ daily registration/verification requests, and utilizing asynchronous workflows, each 2.0 stack handling 50+ steps per request.
• ▪ Deployed automated registration and verification for Amazon Business COVID-19 Supply Store, guaranteeing prompt access to medical supplies for eligible institutions within a two-week timeframe.