2024 – Present: Infrastructure for Large-Scale Recommendation Models

• Architected large-scale ML inference infrastructure for Meta’s recommendation systems, including multi-layer inference caching and disaggregated distributed model serving architectures to improve efficiency and scalability of large recommendation models.

• Built infrastructure enabling recommendation models to consume ultra-long user engagement sequences (10k+ events), including data pipelines, model integration layers, and real-time sequence ingestion used by next-generation recommendation models.

• Partner with multiple ML teams to deploy increasingly complex recommendation models in production, improving system throughput, infrastructure efficiency, and model capability at Meta scale.

2020 – 2023: Recommendation Delivery Infrastructure

• Architected infrastructure enabling in-session real-time recommendations for Facebook Feed, allowing ranking models to react to user interactions during a browsing session.

• Redesigned recommendation delivery and pagination architecture, enabling 2× growth in recommendation volume while significantly reducing serving latency.

• Directed the migration to a shared recommendation serving backend across Facebook and Instagram, enabling multiple product surfaces to operate on a unified infrastructure platform.

2019: Ads Ranking

• Worked on ranking models and feature systems for Click-to-WhatsApp ads, improving advertiser value and model performance.

2016 – 2018: News Feed Delivery

• Built and optimized News Feed content delivery systems, including recommendation modules that surface relevant content when users finish consuming new posts.

Conducted research on Catalan structure shuffling with Dr. Parsad Tetali in School of Mathematics.

• Explored different structures that satisfy Catalan properties (Dyck path, etc)

• Built simulation to estimate the mixing time of different Catalan structures using statistical test

• Experimented with various Catalan structure shuffling techniques

• Visualized the Catalan structure shuffling process with JavaFX

Conducted research on graph mining with Dr. Polo Chau in School of Computational Science and Engineering.

• Visualized graph partitioning algorithms using force-directed layout in D3 and HammerJS, also added support of multi-touch gestures such as pinching to recursively zoom in graph partitions.

• Developed an iOS app that performs real-time personalized movie recommendation, using Swift.

• Implemented scalable graph computation algorithms on Android & iOS and analyzed runtime result. As first author, related work published in IEEE BigData 2014 Workshop.

• Built a new prediction model that determines whether to prefetch comments of a story when the story is just loaded into News Feed on Facebook app on Android.

• Increased number of like, comment & share of all stories by 0.5% globally on Android News Feed.

• led the research that categorizes security events with clustering algorithms

• Implemented and improved Self-Organizing Map algorithm, as well as feature selection techniques and performance metrics that can fit the security event dataset, using scikit-learn package in Python