Xiaofeng Wang

Xiaofeng Wang is an engineering leader at LinkedIn developing the platforms behind its generative AI applications and recruiting agents. Her work spans observational causal inference, production AI infrastructure, and privacy-aware long-term agent memory.

After studying at the University of Virginia, Wang worked in machine learning and data science at Bank of America Merrill Lynch before joining LinkedIn, where she advanced from data science and software engineering into engineering leadership. In 2022, she co-authored the introduction of Ocelot, an internal platform that helps teams estimate the effects of product changes when randomized experiments are impractical. Guided study design, reusable pipelines, automated checks, and expert review made advanced causal methods more accessible without abandoning statistical rigor.

As LinkedIn expanded from GPT-4-powered Collaborative Articles into personalized coaching and Hiring Assistant, Wang helped build its shared generative AI application platform. The architecture reconciled Python-based AI development with existing Java infrastructure and introduced a LangChain-based SDK, versioned prompts, conversational memory, distributed agent orchestration, reusable skills, and OpenTelemetry-based execution tracing. Centralized infrastructure also simplified model switching, governance, and replaying unpredictable agent behavior.

• Agent memory as shared infrastructure. Wang co-developed the Cognitive Memory Agent, which combines conversational, episodic, semantic, and procedural memory. Its ingestion and retrieval layers help agents remember interactions, infer enduring preferences, and adapt to individual working styles while enforcing privacy and application boundaries.

• Privacy-aware hierarchical retrieval. Her co-authored research on Hierarchical Long-Term Semantic Memory, accepted to KDD 2026, describes a production system used by LinkedIn Hiring Assistant. A schema-aligned memory tree organizes information by enterprise ownership boundaries, restricts retrieval to authorized subtrees, supports incremental updates, and preserves source provenance. Wang has identified scalability, latency, retrieval quality, privacy, and operational reliability as essential production constraints.

• Platforms that earn adoption. Wang favors building around immediate developer needs, reusing established infrastructure, and expanding capabilities as products require them. She prioritizes strong software-engineering fundamentals, learning capacity, and teams combining infrastructure, data science, and applied AI expertise.