About

I’m a Ph.D. candidate in Industrial Engineering Technology at Purdue University, expecting to graduate in December 2026. I build reliable LLM agents and decision systems at the intersection of machine learning, optimization, and simulation.

My current research studies tool-using LLM agents for dynamic fleet assignment. At each decision point, an agent examines real-time request and fleet information, then assigns requests to vehicles while balancing service quality and operating efficiency across different spatial and temporal demand patterns. My broader research covers stochastic fleet and battery sizing, routing, large-scale system design, NLP-based social perception, and mobility simulation.

Most recently, I interned as an Applied Scientist at Amazon’s Selling Partner Trust & Store Integrity organization. I developed and deployed a two-stage LLM agent for appeals from sellers restricted by fraud-detection controls. The system combined retrieval-augmented generation (RAG) with graph-based memory to ground recommendations in multimodal evidence, verification actions, similar historical cases, and human corrections. It reached 96.3% cumulative alignment with human experts in production. The work also led to a first-authored arXiv preprint.

I’m open to full-time Applied Scientist, Research Scientist, Research Engineer, and Machine Learning Engineer opportunities beginning in early 2027. You can find my current resume here.

Research Interests

• LLM Agents & Decision Systems — Tool use, RAG with graph-based memory, structured decisions, verification, and fallback
• Optimization Under Uncertainty — Stochastic optimization, online assignment, routing, fleet sizing, and resource allocation
• Simulation & Autonomous Mobility — Dynamic dispatch, pickup-and-delivery systems, delivery robots, EV fleets, and mobility markets

Education

Experience

Amazon — Applied Scientist Intern | Seattle, WA | Aug – Nov 2025

Selling Partner Trust & Store Integrity (TSI) — Fraud Detection Team

• Built and deployed an AI decision-support system for appeals from sellers whose accounts had been restricted by fraud-detection controls, helping investigators distinguish false-positive restrictions from valid enforcement.
• Developed a two-stage LLM agent on Amazon Bedrock using RAG with graph-based memory. Offline, it transformed seller-review records, verification actions, and multimodal evidence into Evidence-Action-Factor-Decision graphs; online, GraphRAG retrieved similar cases to produce evidence-grounded recommendations.
• When evidence was insufficient, the agent requested targeted verification steps and missing evidence. Across 260+ cases, it improved offline alignment from 70.8% to 95.8%, reached 96.3% cumulative production alignment, and maintained 100% precision on approvals.

Selected Papers

1. Evaluating Bounded Language-Model Interventions for Online Assignment in Dynamic Pickup-and-Delivery
   First author — manuscript under review

2. Reviewing the Reviewer: Graph-Enhanced LLMs for E-commerce Appeal Adjudication
   Du, Li, Huang — arXiv preprint, 2026 · arXiv · Details

3. EVC-S: A High-Fidelity Simulation Framework for Three-Sided Electric Mobility Markets
   Vu, Du, Le, Chow — under review at Simulation Modelling Practice and Theory, 2026

4. Two-stage stochastic fleet and battery sizing with routing optimization for sidewalk delivery robots
   Du, Yang, Chow, Le — Transportation Research Part E, 2025 · DOI · Details

5. Analytical model for large-scale design of sidewalk delivery robot systems with bounded makespans
   Yang, Du, Le, Chow — Transportation Research Part C, 2025 · DOI

6. Understanding Social Perceptions, Interactions, and Safety Aspects of Sidewalk Delivery Robots Using Sentiment Analysis
   Du, Le — Transportation Research Record, 2025 · DOI · Details

Beyond Research

Outside research, I like visiting national parks, making pixel art, playing RTS games, mostly Age of Empires IV, and maintaining my Obsidian knowledge system.