Introduction One of the more interesting recent developments in Retrieval-Augmented Generation (RAG) is Knowledge Graph–based RAG. Unlike traditional vector-based retrieval, this approach introduces structured relationships between entities, enabling deeper reasoning and better contextual understanding. This naturally raises a question: Can we build more robust and reliable software systems by combining: Ontologies Domain-specific design models Spec-driven development (SDD) This article explores that idea through a practical comparison.

Overview Running large language models locally is becoming increasingly practical, even for developers without access to massive GPU clusters. In this post, I walk through how to: Run Qwen3.6 35B (A3B, Q4_K_M quantized) locally using Ollama Integrate the model into VS Code Use it as a local coding assistant This setup is especially useful for: Air-gapped environments

Finally, I managed to get vLLM running on my AMD AI MAX+ 395 GPU on Ubuntu 24.04. It was not straightforward — ROCm support on Ryzen AI (gfx1151) is still evolving, and I ran into multiple low-level GPU faults before finding a stable setup. This post documents: - What didn’t work - The errors I encountered - The working configuration Hopefully this saves you a few hours (or days).

Introduction Recently, I’ve been experimenting with a variety of LLM tooling ecosystems, including: Claude Code Codex via OpenRouter Ollama LM Studio vLLM LiteLLM My goal is to better understand the underlying technologies and explore how to operate these tools in air-gapped or controlled environments. In many enterprise settings, developers are allowed to use GitHub Copilot, but not Claude Code.

I first learnt the idea of the Ralph Loop around January, while following developments in LLMs and AI through multiple channels — video feeds, X, newsletters, GitHub repositories, news, and research papers. That sparked a question: What happens if we combine the Ralph Loop with spec-driven design to generate real, working applications? This blog is a reflection of that exploration. My Background with Spec-Kit Since December last year, I have been using Spec-Kit in both:

I have been using GitHub Copilot and Copilot CLI for about one year. During that time, I also experimented with several developer tools and AI-assisted coding technologies. Recently I became interested in exploring Claude Code and related tools such as Claude Cowork. However, I did not want to immediately commit to a subscription before understanding whether Claude Code actually fits my workflow. So I spent some time experimenting with different ways to run Claude Code without paying for a subscription first.

Things are changing incredibly fast, and AI is undeniably the hottest topic right now. There’s no real option to stand still—the only sensible choice is to embrace the change and learn by doing. Recently, OpenClaw has stood out as one of the most exciting personal AI assistant projects. After following the project for several days and spending time reading through its documentation—especially around skills and agents—I decided to try it out myself.

Building with LLMs: How Spec-Kit Changed My Workflow Over the past three years, I’ve used LLMs—such as ChatGPT, Copilot, and Claude—primarily as assistants to generate code snippets for very specific requirements. They were useful, but always in a limited way: small pieces of logic, isolated helpers, or syntax scaffolding. What I didn’t do until recently was build an entire system around an LLM. The main reason was simple: most of my real work lives inside internal frameworks, legacy systems, and opinionated architectures.

Introduction Several days ago, I setup ollama on my MacBook M4, and it works pretty well. At that time, I tried to use it with copilt with local models codegemma:7b and qwen3:8b. My expectation was not so high as the hardware configuration of my macbook pro m4 is just a entry level, just want to see how it works. I also learned there are other options such as vllm. After comparing the two, I found vllm is more flexible, powerful, product-ready, used widely in enterprises.

After evaluating several local Kubernetes solutions I encountered repeated manual steps (downloading bootstrap images, pulling images from registry mirrors and retagging them, and loading images into clusters) that were time-consuming and error-prone. I decided to set up a properly air-gapped Kubernetes cluster using kubeadm and containerd, leveraging containerd’s registry mirror support. This post documents the steps I followed. Prerequisites This guide assumes a Debian/Ubuntu host. Installing the latest Docker Engine will also provide containerd as a dependency.