Weizhi Du

Education

• Research Interests: Foundation models and agentic AI; multimodal learning; retrieval and ranking; AI for health.
• Key Courses: Machine Learning, Computer Vision, Deep Reinforcement Learning, Multi-Agent Systems, Algorithms for Nonlinear Optimization, AI for Health, Analysis of Algorithms, Statistical Data Analysis, Systems Software.

Research Experience

Advisor: Jin Zhang, PhD, MPHS

• Self-evolving RAG: Developed multimodal and multi-agent components of eRT-RAG, integrating query reformulation, curated retrieval, evidence-weighted aggregation, citation grounding, and feedback-driven knowledge updates.
• Evaluated eRT-RAG on 300 ACR TXIT questions across four LLM generations; improved GPT-5 accuracy from 94.0% to 98.0% and achieved 8.2$\times$ higher odds of better physician-rated accuracy on clinical debates.
• Safety-aware self-improving LLM agents: Designed a multi-agent, tool-calling framework for FDG-PET tumor segmentation that iteratively optimized a MONAI SegResNet using bounded, auditable actions.
• Introduced validation-gated improvement with provenance logging and restricted action spaces; achieved 0.813 test-set Dice versus 0.613 for an unrestricted LLM-agent baseline and 0.807 for the MONAI SegResNet baseline.
• Ongoing research and mentorship: Extending agentic methods to brain metastasis imaging and radiotherapy safety checking; mentoring two undergraduate researchers on related projects.

Selected Publications and Research Outputs

1. Weizhi Du, Tingjun Liu, En Wei, Andrew Willett, Matthew Inkman, Xucheng Wang, Thomas Mazur, Julie K. Schwarz, Mengxin Yu, and Jin Zhang. “A Safety-Aware Agentic AI System for Refining FDG-PET Auto-Contouring in Locally Advanced Cervical Cancer.” Accepted poster, 72nd Annual Meeting of the Radiation Research Society (RRS), 2026.
2. Weizhi Du, Andrew Willett, Matthew Inkman, Tingjun Liu, Michael R. Waters, and Jin Zhang. “Generative AI and Large Language Models.” In Generative AI Models and Reinforcement Learning in Radiotherapy Treatment Planning, Chapter 3. CRC Press, forthcoming.
3. Tingjun Liu, Matthew Inkman, Andrew Willett, Xucheng Wang, Weizhi Du, Jessika A. Contreras, Stephanie Markovina, Julian C. Hong, Michael R. Waters, and Jin Zhang. “Evaluation of Self-Evolving Multi-Agent Retrieval-Augmented Generation in Radiation Oncology.” Manuscript under review at JAMA Oncology, 2026.
4. Tingjun Liu, Xucheng Wang, Matthew Inkman, Andrew Willett, Weizhi Du, Julian C. Hong, Michael R. Waters, and Jin Zhang. “Self-Evolving Agentic Retrieval-Augmented Generation Enables High-Accuracy and Confidence-Aware Radiation Oncology Reasoning.” Accepted poster, 72nd Annual Meeting of the Radiation Research Society (RRS), 2026.

Industry Experience

• Developed ML methods for TikTok Search fine-ranking and search-scroll recommendation systems across RoW, EU, and US markets, spanning session representation learning, sparse architecture search, and feature attribution.
• Designed and validated session-level contrastive learning for DeBERTa query/document embeddings, using same-session 5-second plays as positives and other-sessions as negatives; improved stay-time AUC by 0.18 pp versus online baseline.
• Developed parameter-sharing one-shot neural architecture search for sparse embeddings, reducing 600+ online features by 44% with only a 0.04 pp drop in CTR-ranking quality after 10+ days on 24 H100 GPUs.
• Built perturbation-based feature attribution for the same 600+ features using controlled AUC degradation; selected over one-shot NAS, then used to implement feature-dimension allocation and tail-feature masking.

• Built an LLM search agent for Meituan Search on a 770M-user, 15M-merchant platform, using a fine-tuned 7B intent router to dispatch queries to six specialized models and tool-augmented RAG for merchant discovery and Q&A.
• Designed retrieval ablations and DeepEval CI/CD pipelines for real-time merchant/POI retrieval, raising RAG coverage from 49.6% to 72.3% and reducing zero-result rate by 8.2%.
• Developed a post-training pipeline combining Supervised Fine-Tuning on 30K examples with targeted Direct Preference Optimization on curated failure cases, improving dialogue accuracy by 24% and reducing fallback rate by 65%.
• Distilled and quantized a 72B answer-generation LLM into a 7B deployment model, retaining performance within 7% of the teacher and cutting inference latency by 35% through multi-threaded vLLM serving.

Honors and Awards

• Ernest D. Weiss Junior Award for Academic Excellence, Washington University in St. Louis, 2026; sole annual recipient among all juniors across all six CSE majors.
• Howard Nemerov Scholar, Washington University in St. Louis, 2023–2027; recipient of a four-year merit scholarship.
• Dean’s List, Washington University in St. Louis, 2023–2026; recognized every semester enrolled.
• Competition Honors: S.-T. Yau High School Science Awards, Gold Prize in Computer Science, top 1%; Mathematical Contest in Modeling, Honorable Mention, top 15%; Princeton University Physics Competition, 11th place in the U.S.

Teaching Experience

CSE 2407: Data Structures & Algorithms — 6 semesters: TA (1 semester), Recitation/Grading Lead (2 semesters), Head TA (3 semesters); coordinated 30+ TAs serving 200+ students per semester and contributed to curriculum development.
Instructor: Stephen Cole, PhD

CSE 4207: Cloud Computing With Big Data — 1 semester: Head TA.
Instructor: Todd Sproull, PhD

CSE 5104: Data Mining — 2 semesters.
Instructor: Cynthia Ma, PhD

ESE 4270: Financial Mathematics — 1 semester.
Instructor: Vladimir Kurenok, PhD

ESE 4150: Optimization — 1 semester.
Instructor: Shen Zeng, PhD

ESE 4170: Introduction to Machine Learning — 1 semester.
Instructor: Jinsong Zhang, PhD

ESE 3260: Probability and Statistics for Engineering — 1 semester.
Instructor: Jinsong Zhang, PhD

CSE 247R: Data Structures & Algorithms Seminar — 2 semesters.
Instructor: Cynthia Ma, PhD

Selected Projects

• Developed task-conditioned acquisition of hidden agent capabilities using symbolic diagnosis, LLM-based selection, selective disclosure, and persistent memory; evaluated six methods across 576 validated trials, reducing agent contacts by 61% versus random querying and action-schema disclosure by 62% versus Reveal-All.

• Built a training-free visual grounding pipeline combining YOLOv8n region proposals, OpenCLIP crop-text scoring, rule-based spatial reasoning, and interpretable explanations; ablations improved RefCOCO-M Acc@0.5 from 40.6% (CLIP-only) to 62.8% (full system), achieving 0.614 mIoU across 5,598 validation queries.

• Implemented an AlphaZero-style Gomoku pipeline with self-play, replay-buffer learning, policy-value networks, and threat-aware MCTS pruning; completed $8\times8$ connect-5 training in 7 hours versus an unfinished 48-hour unrestricted run, achieving a 60% peak win rate against pure MCTS.

Skills & Interests

Programming: Python, C/C++, Java, JavaScript, SQL, R, MATLAB, Shell

ML Frameworks: PyTorch, TensorFlow, Hugging Face, MONAI, OpenCLIP, scikit-learn

LLM Methods & Tools: LoRA, SFT/DPO, knowledge distillation, quantization, RAG, vLLM, DeepEval

Systems and Tools: Linux, Git, Docker, Kubernetes, Ray, AWS, GCP, FastAPI, Flask, React, Node.js

Natural Languages: English (Fluent), Chinese (Native), Japanese (Conversational)

Personal Interests: Creative writing, backpacking, landscape photography, digital media production, cooking