Skip to main content

NVIDIA RAG Blueprint を AKS に乗せてみた——Helmコマンド一発で動くまでの理解と気づき

· 9 min read
Rintaro Nakahodo
NLP Researcher · Engineer · Creator

Microsoft AI Tour for Partners に参加した際に、NVIDIA提供のハンズオンセッション「RAG Pipeline with Nemotron on AKS」を体験した。MicrosoftパートナーとしてAzureとNVIDIA双方の技術スタックが組み合わさるセッションで、実際にコマンドを叩いてシステムを動かせる内容だった。

Azure上でNVIDIAのRAG Blueprintを動かすハンズオンに参加した。Helmコマンドを数回叩くだけでLLM推論・ベクターDB・ドキュメント取り込みパイプラインが一括起動し、PDFに質問できるUIまで立ち上がる体験は、想像以上に示唆が多かった。

手順を追いながら気づいたことを記録しておく。

Turning Book Knowledge into Copilot Agent Skills — System Design Principles Edition

· 7 min read
Rintaro Nakahodo
NLP Researcher · Engineer · Creator

After reading eda-san's (@eda_sann) article "Turning Book Knowledge into Agent Skills — System Design Principles Edition," I was inspired to build the same thing myself — this time for GitHub Copilot.

While eda-san used Claude Code's Agent Teams, this article takes a different angle:

  • Ported to GitHub Copilot (VS Code Agent Mode)
  • A close look at the structure and format of a Skill
  • A demo of it running in practice

What Happened on npm on March 31, 2026 — Claude Code Source Leak and the axios Hijacking

· 9 min read
Rintaro Nakahodo
NLP Researcher · Engineer · Creator

On March 31, 2026, two security incidents hit the npm ecosystem simultaneously.

The first: Anthropic forgot a single line of configuration and accidentally published the entire source code of Claude Code. The second: the npm account for axios — a package downloaded 100 million times a week — was hijacked and used to distribute malware. The two incidents are completely different in nature, but both shook trust in npm as infrastructure.

What happened on npm on 2026.03.31

Pushing the Blog's Performance to Its Limits — Optimizing Docusaurus PageSpeed

· 7 min read
Rintaro Nakahodo
NLP Researcher · Engineer · Creator

Following the SEO and performance improvements on the portfolio side (previous article), I turned my attention to the Docusaurus blog on the same day. The starting point was Mobile Performance 94 / Accessibility 94 — not bad, but with room to improve. Running PageSpeed Insights repeatedly kept surfacing new problems. Here's the full record.