Building personalized GR product recommendation model and corresponding ML Infra for new Generative retrieval Queue for Dynamic Product Ads.

• Memory Optimization for trie construction. Handle O(Bn) products scale

• Benchmarking CUDA vs triton vs MPHF

• Constrained/Unconstrained prefix decoding.

• End to end delivery in the current Meta's Ads delivery flow across variety of services

Software engg in Ads Monetization org working on delivery, ranking, signal growth.

Ads ranking

• Automation Centric delivery: Enable heirarchical/semantic modeling across delivery and ranking stack.

• Enable multi-stage consistency from retrieval to early to late stage ranking.

Product ranking and retrieval for dynamic ads:

• Improved product ranking late stage ctr and conversion models post ATT signal loss.

Buy with Prime (ecommerce):

• Ranking and delivery lead for Amazon and Meta's collaborative product bwp (https://www.cnbc.com/2023/11/09/meta-lets-amazon-users-buy-on-facebook-instagram-without-leaving-apps.html)

• Lead sev investigator/solver for major ecommerce and shopping Ads performance issues. This required detailed in-depth understanding of entire ads stack (from Ads manager to final Ad impression) across placement types (IG/FB) to identify the underlying root-cause.

• Own P&L for entire trading desk

• Conducted in-depth market analysis, led discussions with stake holders and identified leverage areas, to create a future roadmap for trading desk

• Led a team of 5 members to develop and streamline data pipeline across products and markets

resulting in a trading strategy across multiple trading desks, reducing development time from 6 months to 3 weeks

• Set up trading infrastructure in AWS (EC2, RDS, Lambda, ECR, CI/CD with Git) including but not limited to data pipelines, modeling framework and algorithmic strategy creation.

• Designed and implemented state of the art trading strategy based on probabilistic pricing using quantile random forest on 700+ features, improving profitability by 50%.

• Implemented parallel computing infrastructure using Pub-sub/Docker and Kubernetes with ability to generate one million models per hour.

• Standardized trading strategy backtesting process and KPIs enabling all teams to develop 8 new uncorrelated strategies in 2 years increasing profit by 65% and market coverage by 30%

• Improved maintainability of code by implementing companywide unit tests for key objects. Increased coverage from 0 to 90% in 6 months