FB Reels ML foundation
Leading next generation zero collision hash embedding storage solution in torchrec and FB Reels.
* Used to power ZCH for large cardinality tables (high cardinality).
* Used as cache solution for frozen embeddings.
Leading the cross-org effort to improve GPU training and inference efficiency for all FB Reels model including but not limited to (retrieval, early stage ranking, late stage ranking). Achieved X MW of savings across the board.
Ads ML automation
Worked on blueprints to automate model authoring/experiement and productionalize. This effort is widely adopted by Ads MC and LC process to ship major modeling innovation to all models.
Core system: Data center network routing
Worked on high performance distributed system for routing protocal operated in all FB datacenters.