Mu-En Chiu邱沐恩

A reinforcement-learning dispatch optimization system for Taipei's Dunhua Main Line, designed to reduce bus bunching while balancing passenger waiting time and operating cost.

• Integrated TDX route data, Taipei Bus GPS APIs, and CODIS weather data into a dispatch simulation environment.
• Modeled departure control as a sequential decision problem and trained a PyTorch DDQN agent.
• Balanced bunching risk, passenger waiting time, and operating cost in the reward design.
• Reduced bunching volume by 22.9% and bunching occurrence time by 16.2% versus a fixed-headway baseline.

I used real H&M transaction data as market signals to design a generative AI framework that predicts fashion trends and visualizes potential products.

• Integrated LLaVA and Vicuna to produce trend-aware prompts.
• Used 106 weeks of H&M transaction data.
• Fine-tuned a Latent Diffusion Model with Direct Preference Optimization.
• Improved prediction accuracy by 4-7% over baseline, supporting inventory planning and market response.

I developed an AI pipeline that automates daily news highlight generation through summarization, speech synthesis, and image retrieval.

• Designed an automated news summarization pipeline.
• Combined NLP, text-to-speech, and image retrieval.
• Managed development milestones and module integration.
• Built daily news summarization workflow automation with Python and APIs.

I led a 6-member team to model a beverage chain's operations, identify bottlenecks, and evaluate process improvements through computer simulation.

• Led a 6-member team.
• Built a computer simulation to analyze operational bottlenecks and test improvements.
• Coordinated cross-functional discussions.
• Simulation results showed a 70% reduction in customer waiting time without additional operating cost.