Hi, I'm Xingfeng
ML engineer and researcher. From first-principles materials discovery to billion-scale recommendation systems — building at the intersection of computation, ML, and complex systems.
About
I build ML systems that work at scale. My path has been non-linear: physics at Tsinghua, a PhD in computational physics at Maryland, display ML at Apple, and now ads retrieval and recommendation at Meta.
Along the way, my research on solid-state battery electrolytes has been cited 10,000+ times, and I've shipped patents at Apple and production ML systems at Meta. I'm now drawn to agentic AI and multi-agent systems — working toward AI that can truly understand and change the world.
Experience & Education
AI Research Scientist
Meta
2022 – Present
Ads ranking and recommendation systems at production scale, across two core teams.
- Ads Ranking: Co-authored HSNN, the core ML architecture behind Meta Andromeda — next-gen ads retrieval engine enabling 10,000x model complexity, +6% recall, +8% ads quality across Facebook & Instagram
- RecSys Ranking: Core member of the team integrating LLMs into Threads recommendation, pioneering LLM-powered content understanding and ranking for Meta’s text-first social platform
Display Machine Learning Engineer
Apple
2018 – 2022
Applied ML to display physics, production optimization, and user experience across all Apple products.
- Led ML-based mass production data analysis (PCA, SVM, clustering, anomaly detection), improving display quality and reducing production costs
- Developed deep CNN for mechanical failure mode classification, accelerating display design and testing
- Created display quality metrics combining color science and statistics, adopted across all Apple display products
- 2 granted patents: foldable display compensation (US11817065) and OLED leakage reduction (US12369474)
Ph.D., Physics
University of Maryland, College Park
2013 – 2018
Computational materials science under Prof. Yifei Mo. Research on solid-state battery electrolyte design via first-principles methods and machine learning.
- Designed vector representation of crystal structures enabling ML in materials science; discovered 13 new materials via unsupervised learning
- Pioneered computational studies of fast ion diffusion and solid-solid interface stability
- 17+ publications in Nature Materials, Nature Communications, Joule, etc. with 10,000+ total citations
- 16 issued patents; Co-PI of NSF award 1550423
B.S., Physics (Highest Honors)
Tsinghua University
2009 – 2013
Awards & Leadership
- Advisory board member, Paperclip — largest Chinese popular science video channel, 10M+ subscribers globally (2020–2021)
- Vice President, Kedao — community promoting green energy and public policy, 100+ members across disciplines (2016–2019)
- Vice President, Tsinghua Alumni Association at Greater DC Area (2015–2019)
- Co-PI, NSF Award 1550423 (2017)
- Fellowship and Scholarship, UMD and Tsinghua University (2009–2018)
Research Highlights
HSNN: Hierarchical Retrieval at Meta Scale
Co-authored HSNN, the core ML architecture behind Meta Andromeda — Meta's next-generation personalized ads retrieval engine. Advances beyond the Two Tower model with sublinear O(log N) cost, enabling a 10,000x increase in model complexity, +6% recall improvement, and +8% ads quality on selected segments.
Solid-State Battery Research Pipeline
A computational pipeline from stability prediction to ML-accelerated materials discovery for solid-state battery electrolytes, spanning my PhD and post-doctoral research.