Jun 2022 – Dec 2022 · 6 mos · Palo Alto, California, United States

• Keywords: Trustworthy AI, NLP (Natural Language Processing), Published in Nature Machine Intelligence (2024)
• Conducted a large-scale analysis of 32,000+ AI model cards hosted on Hugging Face to evaluate and improve documentation practices in AI.
• Investigated key aspects of model transparency, including environmental impact, limitations, and evaluation metrics, to promote trustworthy AI development.
• Employed advanced data science and natural language processing techniques to identify community trends, highlight gaps, and propose actionable improvements to industry documentation standards.
• Demonstrated the impact of comprehensive documentation on model adoption and usability, bridging the gap between developers and end-users.
• Collaborated with a multidisciplinary team to publish findings as the first author in Nature Machine Intelligence (2024), establishing a foundational study on AI documentation practices.
• Advanced expertise in responsible AI, transparency frameworks, and ethical deployment of machine learning models.

Jan 2021 – Jan 2021 · 0 mo · Stanford, California, United States · On-site

• Keywords: Medical AI, Computer Vision, Healthcare ML, Dermatology, Model Bias, Health Equity
• Collaborated on development of the Diverse Dermatology Images (DDI) dataset, analyzing 656 pathologically-confirmed dermatological images across Fitzpatrick skin types I-VI (208 FST I-II, 241 FST III-IV, 207 FST V-VI)
• Evaluated performance of 3 state-of-the-art dermatology AI models (ModelDerm, DeepDerm, HAM10000), documenting ROC-AUC drops from 0.92-0.94 on original test sets to 0.56-0.67 on diverse skin tones
• Contributed to model fine-tuning experiments that closed the performance gap between FST I-II and FST V-VI, achieving ROC-AUCs of 0.73-0.77 for light skin and 0.76-0.78 for dark skin
• Analyzed dermatologist consensus labeling accuracy, identifying significant disparities between FST I-II and FST V-VI (sensitivity 0.72 vs 0.59, p < 8.8 × 10⁻⁶)
• Helped create new benchmark dataset enabling systematic evaluation of AI model performance across skin tones, supporting development of more equitable medical AI systems
• Co-authored paper published in Science Advances (2022), providing quantitative analysis of bias in dermatology AI models for improving access to care for 3+ billion people globally lacking dermatological services

Jan 2021 – Jan 2021 · 0 mo · Stanford, California, United States · On-site

• Keywords: Technology Transfer, Data Science, Innovation Analytics, Natural Language Processing
• Led comprehensive computational analysis of Stanford's technology transfer ecosystem, analyzing 4,512 inventions from 6,557 inventors over a 50-year period (1970-2020) to characterize commercialization patterns and success factors
• Developed ML classifiers achieving 0.76 AUROC for predicting commercial success from invention abstracts, identifying key linguistic features correlated with revenue outcomes
• Quantified significant temporal trends including tripling of female inventors (6.5% to 19.7%) and expansion of average team size (2.47 to 3.29 members in biology) from 1995-2020
• Demonstrated that self-licensed inventions through inventor startups generated highest returns - 100% of inventions with >$10M net income were self-licensed vs only 16% for <$10K income inventions
• Identified revenue performance disparities across fields, with top categories shifting from electronics pre-2000 to biology/chemistry post-2000, providing strategic insights for technology transfer offices
• Published findings in Cell Press Patterns (2022), offering first large-scale computational framework for analyzing academic innovation and commercialization patterns

Jun 2020 – Sep 2020 · 3 mos · Sunnyvale, California, United States

• Keywords: Multi-Modal AI, Multi-Modal Transformer, Neural-Symbolic AI
• Developed LRTA (Look, Read, Think, Answer), a novel neural-symbolic reasoning framework for multimodal AI that enhances large language models' ability to reason across text and visual inputs. Advanced state-of-the-art performance by 15% in comprehensive reasoning tasks.
• Key contributions:
• Architected an end-to-end trainable foundation model integrating scene understanding, semantic parsing, and neural execution modules
• Pioneered transparent AI reasoning capabilities by implementing human-readable explanations at each reasoning step
• Created evaluation frameworks to assess and validate model robustness against linguistic perturbations
• Collaborated with Amazon Alexa AI research team on advancing explainable multimodal foundation models
• Research published at NeurIPS 2020 (Knowledge Representation and Reasoning Workshop)
• Amazon Science Page:
• https://www.amazon.science/publications/lrta-a-transparent-neural-symbolic-reasoning-framework-with-modular-supervision-for-visual-question-answering

Sep 2019 – Present · 6 yrs 6 mos · On-site

• CS PhD@Stanford | Multi-Modal LLM, Pretraining & MoE Research

Jul 2019 – Sep 2019 · 2 mos · Beijing, China

Apr 2019 – Jun 2019 · 2 mos · Shanghai City, China

Jan 2018 – Jan 2018 · 0 mo · Urbana-Champaign, Illinois Area · On-site

• Keywords: Computer Architecture, Multi-Modal AI, Deep Learning Accelerator (DLA), Memory and Storage Systems.
• Contributed to DeepStore, a pioneering in-storage neural acceleration system achieving up to 17.7× faster query performance and 78.6× better energy efficiency compared to GPU+SSD systems
• Collaborated on design of novel channel-level neural accelerator architecture optimized for intelligent query workloads under strict SSD power/resource constraints
• Helped develop similarity-based query cache system leveraging neural networks, delivering up to 25.9× performance improvement for intelligent queries
• Participated in comprehensive analysis of storage I/O bottlenecks across visual search, audio matching, and text retrieval applications
• Co-authored paper published at IEEE/ACM International Symposium on Microarchitecture (MICRO '52), a top-tier computer architecture conference