Stochastic Approximation from First Principles
A 22-part series on why noisy iterative updates can converge, from error recursions and step-size conditions to the Robbins–Monro conditions and Dvoretzky's theorem.
AI systems, optimization, and transportation simulation
Yuchen Du
I am a Ph.D. candidate at Purdue University building reliable AI systems for decisions under uncertainty, from graph-enhanced LLM agents to stochastic optimization and simulation models for transportation systems.
Field notes and photosSelected Work
Research
LLM Reasoning
Reviewing the Reviewer
Graph-enhanced LLM agents for e-commerce appeal adjudication, grounded in verifiable Evidence-Action-Factor-Decision graphs.
Stochastic Optimization
Sidewalk Delivery Robot Fleet Design
A two-stage framework for fleet and battery sizing under uncertain delivery demand.
Transportation Analytics
Public Perceptions of Delivery Robots
Sentiment and topic analysis of YouTube discussions around sidewalk delivery robots.
Focus
What I work on
- LLM reasoning systemsKnowledge graphs, action grounding, and auditable decision workflows.
- Optimization under uncertaintyFleet sizing, routing, resource allocation, and stochastic approximation.
- Autonomous transportation systemsSimulation and operational planning for delivery robots and EVs.
Beyond Research
Small signals outside the CV
Outside research, I like visiting national parks, making pixel art, playing RTS games, mostly Age of Empires IV, and maintaining my Obsidian knowledge system.
Browse photosRecent
Writing and updates
Building Auditable LLM Systems
An 11-part series on designing LLM decision systems whose information, tools, actions, memory, and evaluation can be validated and traced.
sdr-sidewalk-sim: Pedestrian-Aware Sidewalk Delivery Robot Simulator
A sidewalk delivery robot simulator that connects OpenStreetMap walk graphs, pedestrian-aware routing costs, dynamic request release, solver-visible snapshots, and replayable visualization artifacts.
A Generative Model Is Like a Database: Decomposition Defines Its Schema
Treating a generative model like a database reveals its missing schema: data must be decomposed into operational units. A Tetris example shows how decomposition granularity sets the model’s inductive bias, compositional freedom, and cost.
We Understand the World by Building Maps
People do not simply apply different rules to identical inputs. Attention, experience, and representation shape the inputs themselves. Science makes maps shared and testable, but those maps remain bounded by instruments, scale, and domain.
Reviewing the Reviewer: Graph-Enhanced LLMs for E-commerce Appeal Adjudication
We transform historical reviewer corrections into retrievable Evidence-Action-Factor-Decision graphs, then use them to ground recommendations for new seller appeals. The deployed system achieved 96.3% cumulative alignment with human experts in production.