Recommendation & personalization
Recommendation systems in production.
Understanding customers and improving the pipeline around the model.
The broader system
Our published work describes personalization as a system of customer understanding, model development, evaluation, and shared engineering infrastructure. Customer feedback and user research inform the modeling choices; offline evaluation and online experiments help assess them.
My contribution to a model improvement
Working with the team, I helped explore a deep learning model to improve on an existing XGBoost baseline, balancing customer preferences and business objectives including GMV and average order value.
I also helped architect the end-to-end setup, from data preparation and model training to deployment.
I helped explore model and architecture choices, GPU training, deployment optimizations, and latency requirements. I focused on understanding which features mattered, improving data preparation and precomputation, and standardizing interfaces for handovers between data scientists and engineers.
What changed
These improvements shortened iteration cycles and time to market, reduced maintenance overhead, and lowered training and inference costs. The model achieved statistically significant improvements in target metrics in an A/B test and is deployed across brands and markets.
Published technical accounts
The articles below describe the wider system and its evolution; they provide context alongside this account of my contribution.