Portfolio

Structured profile

LLM-readable profile

A plain-language rendering of the verified data used by the human-facing portfolio.

Identity and contact

ChunYuan Hsu (許峻源) — AI/backend engineer working across agents, infrastructure, and applied ML.

Master of Science in Data Science student at the University of Michigan building AI agents, backend infrastructure, and applied machine learning systems across production and research environments.

Experience

TSMC — Digital Workflow Development Department Intern

Hsinchu, Taiwan · May 2026 – Aug 2026

Building an AI-agent workflow for structured backend incident investigation across alerts, logs, and metrics.

  • Developed an AI-agent workflow with the Claude Agent SDK to automatically triage backend alerts and generate structured incident-analysis reports.
  • Built integrations to ingest alerts from Alertmanager and retrieve logs and metrics from Kubernetes workloads via ELK and Prometheus.
  • Designed a hypothesis-driven investigation loop that correlates alerts, logs, and metrics to identify likely root causes and summarize actionable findings for engineering teams.

Technologies: Claude Agent SDK, Alertmanager, Kubernetes, Elastic Stack (ELK), Prometheus

QNAP — Backend R&D Internship

Taiwan · Jan 2025 – Jul 2025

Built retrieval and developer-tooling systems, improved a Go service, and diagnosed production reliability problems.

  • Built a retrieval-augmented Jira issue search system using AWS Bedrock and ChromaDB embeddings, increasing developer issue-resolution efficiency by 50%.
  • Developed an MCP-based Jira search server that integrates with IDEs, enabling developers to query and explore issues directly from their coding workflow.
  • Refactored Device Avatar microservices from Python to Go, achieving a 30% performance gain and optimizing deployment on Kubernetes.
  • Diagnosed and patched a critical memory leak in cloud production by correlating Grafana metrics with execution traces.

Secondary verified detail:

  • Migrated the Konnyaku service from Python 2 to Python 3 and deployed it on Kubernetes.
  • Added token authentication and unit tests to Device Avatar.
  • Benchmarked MongoDB and Couchbase for service storage.
  • Investigated DDNS worker failures during RabbitMQ scaling.
  • Investigated NATS connection failures under production scaling.

Technologies: AWS Bedrock, ChromaDB, MCP, Go, Kubernetes, Grafana

Lilac

Context: A broader cross-cloud Infrastructure-as-Code lifting research system that learns reusable mappings from deployed cloud state to Terraform and uses LLM assistance with symbolic and Terraform-native verification.

My contribution:

  • Built a focused graph-based lifting workflow for concrete Terraform mappings, primarily on Azure, including JSON-schema and dependency cases.
  • Implemented Azure cloud-state collection and a GCP resolver that uses an LLM to infer service-specific CLI commands, then parses and caches them by asset type.

Project result:

  • The work established an implementation-oriented foundation for Lilac’s broader cloud-agnostic, correctness-aware lifting pipeline.

Technologies: Terraform, Python, Large Language Models, Symbolic Verification, Azure, Google Cloud

Evidence links:

Toward Interpretable Brain Age Prediction and AD Classification

Context: A University of Michigan EECS 545 team project for brain-age regression and Alzheimer’s Disease classification using structural MRI from OpenBHB and ADNI.

My contribution:

  • Focused on infrastructure and data processing, including generation of patch-level 3D embeddings and coordinates used by the downstream models.

Project result:

  • 0.873 diagnostic accuracy
  • 0.775 macro F1
  • 3.54-year MAE
  • 0.966 R²

Technologies: PyTorch, NeuroVFM, 3D MRI, Multiple Instance Learning

Evidence links:

VizThinker

Context: A graph-based interface for interacting with LLMs that replaces a single linear transcript with a visual conversation graph.

My contribution:

  • Implemented branching and node-based history navigation for complex idea exploration using Node.js, React, and Python.
  • Deployed the application on Google Cloud Platform.

Project result:

  • The project produced a graph interface for branching through conversation history.

Technologies: Node.js, React, Python, Google Cloud

Evidence links:

Additional projects

  • Jira Issue Search (QNAP internship work): A repository for the retrieval-augmented Jira search work, using AWS Bedrock and ChromaDB; the internship outcome and metric are stated in Experience. Technologies: Python, AWS Bedrock, ChromaDB View source
  • Issue Search MCP (QNAP internship work): An MCP server that exposes natural-language Jira query, suggestion, and issue-retrieval tools to coding workflows. Technologies: Python, MCP View source
  • File Translator (Independent project): A Gemini-powered tool that translates English PDF documents into Traditional Chinese while preserving layout through generated LaTeX. Technologies: Python, Gemini, LaTeX View source
  • AZtec Image Comparison (Computer vision project): A computer-vision tool for detecting and comparing overlapping patterns in crystallographic pole-figure images. Technologies: Python, OpenCV, NumPy View source
  • MIPS CPU Architecture (Computer architecture coursework): Verilog coursework covering MIPS assembly, an ALU, a single-cycle CPU, and a pipelined CPU with forwarding and stalling. Technologies: Verilog, MIPS View source
  • OS Nachos (Operating systems coursework): Operating-systems coursework implementing system calls, multiprogramming, virtual memory, and file systems in Nachos. Technologies: C++, Nachos View source
  • Advanced Compiler (Compiler coursework): LLVM coursework implementing data-dependency and pointer-analysis passes and studying array languages. Technologies: C++, LLVM View source
  • Quantum Event Identification and Simulation of Quantum Event-Learning Procedures (NTHU thesis and research project): Python simulations comparing quantum random and blended measurements for quantum event identification. Technologies: Python, Quantum Simulation Read report; View source

Skills

  • Languages: Python, C++, Go
  • AI & ML: PyTorch, AI Agents, Claude Agent SDK, LLM Integration (AWS Bedrock), RAG (ChromaDB), Embeddings, Semantic Search
  • Cloud & DevOps: Docker, Kubernetes, AWS, GCP, GitLab CI/CD, Prometheus, Elastic Stack (ELK), Alertmanager, Grafana, NATS
  • Frameworks & Systems: Node.js, React, Linux (Debian, Arch), Git, Scrum

Education

  • Master of Science in Data Science, University of Michigan, Ann Arbor, MI, USA — Sep 2025 – Present
  • Master of Science in Computer Science, National Tsing Hua University, Hsinchu, Taiwan — Sep 2022 – Jan 2025
  • Bachelor of Science in Civil Engineering, National Cheng Kung University, Tainan, Taiwan — Sep 2018 – Jun 2022

Earlier roles

  • Volunteer, US Taiwan Watch — 2024: Developed backend features for the organization’s website.
  • Teaching Assistant, Linear Algebra — 2023–2024: Supported international students in mastering Linear Algebra concepts.

Records

Personal background

Originally from Tainan, Taiwan, ChunYuan Hsu’s path followed civil engineering to computer science. Outside technical work, his interests include watching sports, especially baseball, basketball, and football, going to the gym, playing darts, Linux ricing and interface customization.