Sean (Yusheng) Han

MS, Information Systems

• Coursework: Prompt Engineering & AI, Program Structure & Algorithms, Application Engineer & Development, Data Science Engineering Methods, Data Management & Database Design

• Engineered a HIPAA-compliant, AI-powered voice agent capable of conducting complex phone surveys in multiple languages with ElevenLabs; successfully handled real-world call interruptions with a 99% survey completion rate.
• Designed a multi-agent AI workflow with LangGraph that automatically extracted and structured data from thousands of unstructured patient surveys; accelerated data processing by over 500% and generated insights that led to a 15% improvement in consultant services.
• Developed a secure full-stack web portal with FastAPI and React.js for patients and consultees using JWT authentication; architected to handle 1000+ concurrent users with RabbitMQ, delivering sub-2-second page load times with zero security breaches.
• Implemented a complete CI/CD pipeline from scratch using GitHub Actions, automating build, testing, and deployment processes; cut manual deployment effort by 95%.
• Architected a scalable HIPAA-compliant database solution on AWS to securely manage sensitive health information for over 1,000 patients while ensuring 100% data integrity and availability.

• Fine-tuned a 1.9B-parameter GPT-style model on an 8x NVIDIA H100 cluster using BF16 mixed precision and fused optimizers, achieving approximately 108K tokens/sec throughput.
• Implemented custom ZeRO-style state and gradient sharding that reduced GPU memory overhead by approximately 7x compared to standard Distributed Data Parallel training, enabling larger batch sizes and better utilization.
• Architected a high-throughput distributed data loader with manual sharding, masked loss computation, and task-aware batching to support multi-task training across terabyte-scale datasets.
• Deployed a hybrid optimizer workflow combining Muon and AdamW with a GRPO-based reinforcement alignment strategy, improving reasoning accuracy in coding-focused benchmarks.
• Identified and resolved NCCL communication bottlenecks and data loader inefficiencies, achieving near-linear scaling efficiency across nodes.
• Built an automated evaluation harness for LeetCode and HumanEval benchmarking with sandboxed execution, enabling secure Pass@1 scoring and reproducible model comparisons.