Browse Stanford University | Stanford Profiles

Xinnan Wang

Professor of Neurosurgery

Current Research and Scholarly InterestsMechanisms underlying mitochondrial dynamics and function, and their implications in neurological disorders.

BioI'm an ML engineer specializing in LLM post-training and agentic systems—with a particular focus on domains where rigor matters: health, biology, and scientific discovery.

Long-horizon agents — Designed and shipped multi-step orchestration systems (Pantheon-CLI, OmicVerse Agent) that outperform general SWE-agent baselines on biomedical tasks. Built cross-provider query routing and sandboxed execution to keep complex workflows robust over extended interactions. My agents don't just respond—they plan, recover from failure, and complete real research pipelines end-to-end.

Agentic science — Created infrastructure where AI doesn't assist research—it conducts it. Vectorized 30 years of NHANES data; parallelized Bayesian kernel machine regression on Kubernetes; built TCGA/GEO pipelines that bridge wet-lab and dry-lab workflows. Co-developed OmicVerse, an open-source platform powering reproducible multi-omics and single-cell analyses across hundreds of studies.

Experience engineering — Scaled rubric-based reward datasets to 1M+ pairs; trained summary and chain-of-thought reward models via RLAIF/RLHF; delivered measurable benchmark lifts in health AI. I care about the full loop: data curation → reward shaping → careful ablation → verifiable outcome—no cherry-picked demos—just metrics that survive scrutiny.

Currently pursuing advanced agentic studies at Karolinska Institutet and Stanford!

Open-source: OmicVerse · Pantheon-CLI · RAG Web UI · AstrBot

If you're working on post-training at scale, scientific agents, or high-integrity data pipelines—I'm always interested in systems that move from promising results to verifiable outcomes. Let's talk.