Greater Boston
I am a data science engineer with 8+ years of hands-on experience in e-commerce and sports betting industries. Feel free to reach out to me at [email protected] .
Building, optimizing, and executing market making algorithms on prediction markets.
At Mojo, I architected and lead technical direction for the ML Systems and Platforms that powered the company's Athlete Stock Market. Some of my initiatives included: - Designed and implemented our real-time inference pipelines for predicting performance for NFL, MLB, NBA, and NCAA athletes. I built these pipelines using Python, Terraform, AWS Lambda, AWS Elastic Container Service, MLFlow, AWS Sagemaker Endpoints, AWS S3, AWS DynamoDB, and AWS Kinesis Data Streams. These pipelines processed around 200k on-field events per-game, and produced a projection update after each event in under 300 milliseconds. The projections produced by these pipelines managed the pricing of thousands of our athlete stocks that take in around $1M in wagers per month. - Architected and built our CI/CD workflows for delivering new models, data services, data migrations, and infrastructure to production. These workflows were built using Bash, Python, SQL, Terraform, MLFlow, AWS RDS, AWS S3, AWS ECR, and GitHub Actions. These workflows supported the velocity of a 15-person data team, and powered the production deployment of 2 new models per week, 10 new data services per week, 5 new infrastructure projects per week, and around 300 merged PRs a month. - Rolled out in-game and post-game monitoring systems for all model inputs/outputs in production. The in-game monitoring system was implemented using Python, AWS Lambda, AWS DynamoDB, and Django, and the post-game monitoring system is implemented using Python, AWS Lambda, AWS Kinesis Firehose, AWS S3, and AWS Glue. These solutions enabled our team to detect feature/model drift phenomena and take actions that saved Mojo around $10k a month in hold. - Mentored junior and mid-level engineers on topics such as systems design, infrastructure-as-code, performant Python, and effective software design.
As the first Machine Learning Engineer at Vroom, I Led ML Platforming initiatives and architected the ML Infrastructure that power our car-buying experience. Some of my projects included: - Upgrading our model services to autoscaled endpoints using Kubernetes, Helm Charts, Terraform, AWS EKS, and AWS ECR. These autoscaled endpoints enabled us to perform segmentation in real-time on all users who activated on the website. This scale of segmentation enabled us to develop personalization features that increased checkout conversion by 5%. - Building inference delivery pipelines that enabled use cases for our model predictions in our search backend and our marketing campaigns. These inference pipelines were built using Python, Terraform, AWS Lambda, AWS ECR, AWS S3, and Kafka. These pipelines powered front-end vehicle sort algorithms and social retargeting ad campaigns that contributed to a 1-week reduction in the average time it took to sell a vehicle. - Migrating our model source control and deployment workflows from Bitbucket to Gitlab. This required developing CI/CD pipelines using Bash, Gitlab CI/CD, Docker, AWS ECR. Completing this migration led to more flexible model deployment workflows that reduced average model productionization time from 3 months to 3 weeks.
As a data scientist at Vroom, I designed and developed our user segmentation systems to help Vroom identify high-intent customers, personalize customer experiences, and establish sales strategies that created a high deposit rate. Some of these initiatives included: - Building a lead-to-contract algorithm that used web interaction data to identify high-intent customers. The algorithm was built using Python, FastAPI, Scikit-Learn, and AWS DynamoDB. The sales strategies powered by this algorithm generated an additional 1000 deposits per year. - Architecting our Customer Activity Tracking System that collected, computed, and served user intent signals for web personalization. This system was developed using Python, Terraform, AWS Lambda, AWS DynamoDB, and AWS API Gateway. This system powered personalization features that increased our checkout conversion rate by 10%. - Mentoring junior data scientists on projects related to model productionization, recommendation sytems, and sort.
As a data scientist at CarGurus, I worked on optimizing the car listings platform by developing products for car recommendation, dealer retention, and search engine marketing. Some of my key projects included: • Developing a neighborhood method for recommending millions of cars on CarGurus’ search pages using Python, PySpark, Snowflake, and AWS EC2. We plan to deploy this recommendation method on search pages in Q2 2020. • Building a model that predicts whether a dealer will cancel their paid service in the next 3 months using Python, LightGBM, SHAP, LIME, and Snowflake. This model has an 82% cross-validated AUC, and it will be used by the sales team to plan dealer retention strategies for each fiscal quarter. • Implementing an algorithm to discover millions of new keywords to bid on in Google Search Engine Marketing using Python, FastText, SpaCy, and Tensorflow.