• Built Agent harness powering all Business AI products.

• Post-training for production agents: synthetic-data SFT, RL, LLM-as-judge

• Invented Position Interpolation — first method to extend LLM context windows up to 16x without architectural changes; Used by Qwen/Gemma.

• Tech lead across 3 products and 150+ engineers

• Invented core modeling architectures that significantly improved model accuracy and scaling, adopted company wide. Author of more than 2% Ads revenue improvement from modeling techniques alone.

• Pioneered Meta's first large-scale Transformer-based recommendation architecture.

• Built key ML infrastructure used by 500+ engineers for 8 years, enabling faster model iteration across Meta's ranking systems. Built efficient ML systems that allow combining hundreds of concurrent modeling works.

• Tech lead of 50+ engineers; set the multi-year architecture research agenda, developing and shipping high-impact modeling ideas

• Built large-scale metric aggregation system for Bing on Spark (up to 10x throughput)