Building a Home AI Stack with the Help of AI

This is the fourth post in this series. After moving more than ten years of notes and attachments (parts 1 and 2) and the blog (part 3) onto my own hardware, this time I tried to bring a newer tool home too: local AI.

Unlike the previous posts, this was not something that had sat on my to-do list for years. It was an unplanned project. I wanted to make full use of a mini PC I had just bought, so I planned to build an AI chat interface that uses my own notes as its data.

What it actually does

When I started planning, my only contact with AI was the online chat service. So the first step was to use AI to learn what AI is, and what a local AI can do.

Before I had looked into it enough, the plan included an AI chat interface, a RAG (retrieval-augmented generation) setup and AI image generation. After a few trials I understood the hardware limits of my mini PC: even a simple image prompt took more than 20 minutes to process. So the goal changed to a platform for simple, repetitive information processing.

After more than half a year of improvements, the platform does two main things: (1) it collects and pulls together investment news every day, to help with my day-to-day choice of stocks and ETFs; (2) it brings my years of notes back to life, with a RAG that I can “talk” to through an AI chat interface. The last step was a frontend that connects to the two backend systems.

What I actually built

The setup includes:

• ollama: runs the AI models (the mini PC uses an AMD GPU, which caused quite a few difficulties along the way).
• open-webui: a chat interface in the browser, for talking to the local models (I rarely use it now that I have my own frontend).
• RAG pipeline: a PostgreSQL database, so the AI models answer questions “based on” my notes.
• backend: a backend system that uses AI to collect, organise and give feedback on investment data (using mathematical models I found online).
• frontend: an interface for “talking” to the backend systems easily.

Using AI to build AI

I have to admit I knew nothing about AI or coding. I did not have the ability to build this system myself. AI played a very important role in the process: the early fact-finding and the design of the platform’s features, and the structure and debugging of the whole backend and frontend. My role was to say what I needed and to oversee the whole process.

While building the system, my understanding of AI kept being updated, and worries about using AI came up from time to time. AI is very powerful, but the problems that come with it cannot be ignored: privacy, network security, and how accurate AI-written code is.

Cloud AI is stronger. Why run it locally?

In fact, cloud models are very good and are developing very fast. In only about half a year (the first half of 2026) they went from mainly offering chat interfaces to vibe coding becoming common. I relied on these tools to finish the earlier migrations too. A local AI running on my mini PC will never catch up with these cloud models in what it can do, but it has its own role in how I use AI every day:

• Privacy: the documents in the RAG store and the financial data it collects stay only on my hardware. That fits the theme of the whole series: my files in my own hands.
• “Free”: I still pay for the electricity, but there is no per-token charge while it runs. It works well for repeated, continuous data collection.

So the split between cloud and local AI is: cloud AI for computing power, local AI for private and scheduled repetitive work.

What I got out of building local AI

With high electricity costs, and a platform that needs regular updates and maintenance, setting up and running local AI is not completely “free”. But the result is not only a service platform. It also let me take my first step towards understanding AI. If I had not set out to build local AI, I would not have started learning what Docker, a context window or RAG are. Building local AI was an investment and a learning process, and that is what I got out of this home build.