May 2025

Highlight of the Month

Summarize my biggest breakthrough, project, or insight in this month:

This month I mainly focused on my side project Fraud Detection: From DataOps to MLOps, and successfully integrated lots of interesting tools I've never tried before, like Feast, MLflow, Ray and KServe! It helped me better understand how to connect data pipelines with real-time model serving and got so much fun during the process.

What I Consumed

A list of articles, papers, courses, or videos I read/watched/completed:

Read

• Real-time Fraud Detection on GCP with Feast
• Hyperparameter Tuning with MLflow and Optuna
• Deploy MLflow models with InferenceService
• Running Tune experiments with Optuna
• ihower's Facebook Post about MCP
• Python Tooling at Scale: LlamaIndex’s Monorepo Overhaul
• Using uv in Docker
• Docker 教學：用 Multi-stage build 建立 Poetry 虛擬環境
• Python 套件管理器——Poetry 完全入門指南

Watched

• Claude 4 ADVANCED AI Coding: How I PARALLELIZE Claude Code with Git Worktrees
• Code with Claude Opening Keynote

Completed Courses

• GitHub Copilot 協作開發實戰
• MLflow in Action - Master the art of MLOps using MLflow tool
• Real-world End to End Machine Learning Ops on Google Cloud

What I Created or Tried

What I built, experimented with, or wrote:

• Set up an end-to-end Data2MLOps pipeline using dbt, Feast, MLflow, Ray and KServe.
• Published a series of posts on my side project Fraud Detection: From DataOps to MLOps.
• Tried out multi-stage Docker builds with uv to optimize my Python environment setup.
• Experimented with MinIO as an S3-compatible object store on Kubernetes.
• Tried out GitHub Copilot's configuraiton best practices and its agent mode for more efficient coding.

What I Learned

Short reflections on what I actually learned or became more confident in:

• Grasped how multi-stage Docker builds work and how they can significantly reduce image size and build time
• Deepened my understanding of how Ray Tune integrates with Optuna to perform distributed hyperparameter tuning, enhancing both speed and efficiency in machine learning tasks
• Learned the pros and cons between REST API and gRPC for model serving, and how to use KServe to deploy models with both protocols

Reflections - Beyond Just Tech

Soft-skill insights or workflow/communication/process reflections:

• Recognized that in the AI era, mastering individual tools is easy, but engineers add value by excelling in multi-tool integration and architectural design.
• Realized that relying solely on LLMs for poorly documented tools can lead to inefficiencies. It's crucial to combine LLM assistance with thorough manual exploration and testing.
• Noticed that GitHub Copilot can significantly speed up coding, but it requires careful management to avoid code quality issues.

Goals for Next Month

Set 2–3 simple goals to stay focused and accountable:

• Explore Airflow 3.0 and its new features.
• Try building my first MCP server.
• Accelarate my developer experience with GitHub Copilot and its agent mode.
• Write and publish my thoughts on why I use mkdocs-material for my blog and how I set it up.