Spearheaded the end-to-end modernization of a core music retrieval system, transitioning from legacy heuristics to a state-of-the-art Two-Tower deep learning architecture, resulting in a significant double-digit lift in key user engagement metrics.

Played a key role in the research and development of Spotify Home's Transformer-based sequential models for recommendation. Contributed to the full model lifecycle.

Architected and implemented scalable MLOps and data infrastructure to support large-scale model development. Key technologies include distributed data processing (Scio), modular training pipelines for rapid experimentation (Tensorflow, PyTorch, Ray, Flyte), and low-latency online serving with vector databases (ANN/HNSW).

Successfully executed multiple production model launches across different recommendation surfaces (e.g., albums, playlists), demonstrating the flexibility and real-world impact of the new deep learning systems.

Gained hands-on experience within one of the world's largest recommendation ecosystems, focusing on the optimization of deep learning models for short-form video (Reels) at a billion-item scale.

Developed a deep, practical understanding of Meta's state-of-the-art ML infrastructure, including production-level data pipelines, large-scale feature engineering, and the application of industry-leading architectures like DLRM.

Designed and pioneered a configuration-driven Apache Spark framework for temporal feature engineering, which became the foundation for the open-source project, Feathr. This innovation drastically reduced the feature development lifecycle from months to just one week.

Engineered and launched novel long-term and short-term user interest features that became a top predictive signal in the newsfeed ranking model. This work delivered a ~2% lift in core business metrics, single-handedly accounting for nearly half of the team's annual OKR.

Led a major initiative to unify the separate desktop and mobile newsfeed ranking models into a single, cohesive architecture. This strategic project streamlined the ML development process and delivered one of the largest annual improvements to key business metrics.

Architected and built a foundational ETL pipeline (the "Feed Master Table") using Spark to reliably process over 15TB of data daily. This centralized data source significantly improved data quality and developer productivity for the entire newsfeed machine learning and data science organization.