Scaled FB Reels Overlay Ads from 0 to $1 B/year through Ads Personalization

Responsible for building ML solutions and designing systems to deploy them.

Built menu digitization system by stacking Amazon Textract with custom Object Detection and Text Classification models. System is being used in production to add advanced search functionality to a restaurant aggregation and food delivery platform.

Developed statistical technique to predict winners of a tournament-style autonomous racing league. Designed architecture to serve results to commentators in real time.

Publication: https://aws.amazon.com/blogs/machine-learning/delivering-real-time-racing-analytics-using-machine-learning/

Implemented architecture to calculate Formula 1 driver rankings.

Publication: https://aws.amazon.com/blogs/machine-learning/the-fastest-driver-in-formula-1/

Financial services consulting. Reduced operating expenses of major US credit card issuer by $81 MM annually. Built machine learning models to: 1. reduce losses with an improved algorithmic credit policy, and 2. optimize contact strategy on delinquent accounts. Navigated stringent regulatory landscape to ensure models were legally compliant.

Created account and portfolio-level valuation forecasts by integrating machine learning predictions with data on customer behavior, seasonality, and macro-economic trends.

Improved runtime of client’s daily reporting code from 12 hours to 60 seconds. Leveraged optimized data structures and array programming in Python to develop solution within client’s production environment.

Designed company-wide training sessions and initialized Pyspark codebase for performing distributed data wrangling and model training on AWS EMR/Sagemaker. Represented firm at AWS re:Invent 2018.

Identified 5 high-confidence candidate mutations for coronary artery disease using machine learning. Accelerated pathology research by providing alternative to wet lab intensive workload of studying mutations correlated with complex diseases.

Trained support vector machines on genomic data to identify harmful mutations. Deployed models on SLURM cluster to achieve scalability when performing tissue-specific inference on genome-scale datasets.

Developed novel approach towards personalized medicine by creating a tool to study the impact of bacterial communities within the body (microbiome) on human health. Outperformed status quo technique (Flux Balance Analysis) on benchmark dataset by ~20%.

Implemented Skip-gram in Tensorflow to create vector representations of bacterial metabolic network reconstructions, method was generalizable to bacterial communities.

Used constraints based metabolic reconstruction models to investigate the relationship between P. aeruginosa metabolic phenotype and pathogenicity/environmental niche.