Jun Peng says 👋

This is a blog for him to build and share in public.

I’m a Data Architect and a Builder with a broad background in business, data science, and data engineering.

My work spans building Golang backend systems, PySpark data pipelines, Data mining, and developer tooling. I have built many 0-to-1 products, and love how AI empowers me to learn and build faster.

Most personal sites don’t need a design system. Mine probably doesn’t either. But I wanted to find out how precisely AI tools like Google Stitch could translate a vague aesthetic into something concrete and codeable.

What I found

Stitch’s DESIGN.md revealed how unusually specific words like tectonics, terracotta, and Intentional Asymmetry is what actually steers the model toward a coherent visual identity.

There were lots of learnings from my further investigations into OKLCH color spaces and Design Token hierarchies.

They say AI agents will finally free us from oncall and let us “touch grass.” I put that to the test. What started as a simple DNS setup evolved into a full K3s cluster featuring Gitea, Grafana, and 1Password Secret Ops.

The Good and Bad

There were surprises like one-shotting the docker compose for the DNS setup, and Gemini’s ability to simulate conversations between 2 experts to come up with a secure and practical home setup.

But there were equally many frustrating hallucinated helm chart configs. Claude still tells you to use deprecated Bitnami charts!