Victor Steinborn

How to Setup Your Personal Home Server - For Scientists and Engineers

2026-07-05T00:00:00+00:00

"Those who know nothing of foreign languages know nothing of their own."
-- Johann Wolfgang von Goethe </blockquote>
Getting Computers to Communicate Simply</a></h1>
I was recently interested in setting up a home server so that I can host my own large language models using Ramalama</a> on an older computer with a dedicated graphics card.
However, when I started learning more about what it takes to set up such a server, I quickly realized a lot of the solutions out there relied on using 3rd party apps or software.
I personally prefer to keep my setup as simple as possible using system defaults and pre-installed software. Ultimately I would like the setup to feel like a computing cluster that can be found at a university for scientific and engineering applications with some websites that people can access from the browser (like one website for chatting with a local LLM).
Additionally, I also like to apply scientific experimental principles to my setup. Specifically, the most important principles for me are reproducibility and simplicity. When running the server, I want my setup to be easy to replicate on another system. Additionally, I don't want to be dependent on an excessive amount of 3rd party software as that also impacts reproducibility and simplicity when the software I depend on requires an update that introduces breaking changes.
This article does not go into great detail on how to get everything running, but instead is intended to provide an outline with links to free resources that provide step-by-step instructions.
Step 1 - Get a Server Running in Your Home Network</a></h1>
This was one of the steps that conceptually confused me the most when first starting out. Many of the guides on getting home servers running usually focus on one specific piece of hardware (e.g. a Rasberry Pi) and getting one specific piece of software running on it (e.g. Nextcloud).
What confused me at the time was the guides made it seem like a home 'server' is a very specific setup, however a server is quite a general concept. In short, paraphrasing Wikipedia</a>, a server is simply a computer that can provide data, resources or services to other computers in a network. This means almost any computer can potentially be configured to act as a server.
</div>

This may be obvious, however it is important to be aware that a server is ultimately just another computer for our purposes. </div> </div> </div>
For our setup, taking inspiration from computing clusters that may be found at universities, I ultimately decided to setup my server using Fedora Server</a> after looking through what options exist from the popular Linux distribution providers. You can also use other operating systems or solutions, however for my use case it was simplest to go with Fedora Server.
To set up Fedora Server on my system I followed this excellent video guide</a> from TechHut. This guide also explains how to get some simple services running using Podman</a> (an alternative software to Docker), which comes pre-installed with Fedora Server.
Additionally as a side note, using Ramalama you can easily create Podman/Docker Compose files for serving LLMs on your server via the --generate=compose</code> flag.
For example, you can create a compose file for running smollm using the following command from the Ramalama documentation:
ramalama serve --name=my-smollm-server --port 1234 --generate=compose smollm:135m </code></pre>
Step 2 - Setting up an NGINX Reverse Proxy</a></h1>
At this point you might already be happy with your server. You would need to open up a port in your Server's firewall for each service you want to make available in your home network. To avoid opening up more ports you can make use of an NGINX reverse proxy that would route traffic from a single port to different services.
For this I benefited from watching Cameron McKenzie's video guide</a> on how to setup an NGINX reverse proxy. Many guides I have seen on this topic make use of 3rd party software that manage the reverse proxy for you, however this guide makes use of docker and uses the base NGINX image. To follow this guide with Podman you would simply need to replace docker</code> with podman</code> in each command.
One thing I would like to note is that I have seen the term 'reverse proxy' used in some guides to describe a setup where you want computers outside of your home network (i.e. computers not directly connected to your router) to access your server. The video does not setup your server for that type of configuration. Here we want the reverse proxy to route traffic coming from within the home network to a specific service running on our server (such as a site hosting an LLM). The service would not be accessible outside of our home network.
Step 3 - Setup Self-Signed SSL Certificates</a></h1> At this point you have a server that has services accessible through a single port, however you might have noticed that if you try to access these services you have a warning in your browser that the connection is not secure. If you ignore the warning you can access your service. However, the only issue with this approach is that anyone/ any device in your home network can read anything you send over this connection because you are sending data over the HTTP protocol. To fix this and use the more secure HTTPS protocol (the S stands for 'secure') you can use self-signed SSL certificates. There are also other ways to tackle this issue, however this worked for my use case and it did not require me to install extra software or pay for a domain name. For learning about the theory behind SSL certificates and for a guide on getting HTTPS setup between devices in the home network, I greatly benefited from Christian Lempa's video guide</a>. Closing Thoughts</a></h1> Setting up my own home server was a rewarding experience for getting practical experience in computer networking and deploying services. Additionally, I found setting up a home server was a great way to take steps to save money by opting for self-hosted services (as opposed to cloud solutions) and taking control of personal data. There are many ways you can setup your server so I hope that you benefited from this guide by helping you decide what setup is best for your use case. @misc{2026-vsteinborn-serverhowto, author = {Victor Steinborn}, title = {How to Setup Your Personal Home Server - For Scientists and Engineers}, year = {2026}, url = {https://vsteinborn.github.io/posts/personal-server/} } </code></pre>

How to Use Ramalama to run Containerized LLMs Privately in Zed 2026-02-17T00:00:00+00:00 "Kizuri chajiuza, kibaya chajitembeza." "Good things sell themselves, bad things need advertising" -- Swahili Proverb </blockquote> Motivation</a></h1> It seems like not a day can go by where we don't hear about AI on the internet lately. Most of the discourse regarding AI focuses on specific tools from providers that you can access by visiting their website. However, these tools are often not free or require you sending your text conversation or other data to the provider to make use of the tool. This last point is especially problematic if you work in a setting where personal data or intellectual property cannot (or should not) be transferred out of an organization. A solution to this dilemma is to run the AI model on your own computer using open-source models. There are multiple ways of doing this, in this post I will show how you can use Ramalama</a> to run models in OCI containers. </div> OCI containers are virtualised operating systems that work in a manner that is standardized between devices. </div> </div> </div> The main motivation for using OCI containers is that you don't have to worry about changing configurations if you decide to move to another device or cloud provider, this keeps the setup clean and reproducible. This post is intended to outline some lessons learned in getting Ramalama running so that you can avoid the same pitfalls I had. Advantages</a></h2> Ramalama and open models are free and open-source</li> You keep your data local and private</li> Open-source models and software can be modified and audited</li> The models are containerized for portability and security</li> </ul> Things to keep in mind</a></h2> Some models are quite large and require a lot of memory</li> Some models (especially large models) can consume a lot of electricity to run</li> </ul> </div> If the model requires more memory than your system is able to provide, you unfortunately can't run the model without changing your hardware or how the model is loaded into memory. </div> </div> </div> Notes on Setup</a></h1> I will specifically focus on running Ramalama (version 0.16.0) for text completion in the Zed IDE</a> (version 0.223.3), however you can use Ramalama for other applications as well. For this setup I am using a computer running Fedora Linux</a> with the default dnf</code> package manager, an NVIDIA graphics card and with Podman installed. This post will not be discussing how to setup agentic features or tools. </blockquote> Setup</a></h1> Install Ramalama</a></h2> I started off with installing Ramalama using the information provided in the official GitHub Repository</a>. To do this I ran: sudo dnf install ramalama </code></pre> Install NVIDIA Container Toolkit</a></h2> Subsequently, because I am using an NVIDIA graphics card, I had to install the nvidia-container-toolkit</code> to enable CUDA support in containers. This is where I ran into my first issue, because no nvidia-container-toolkit</code> package exists in the default Fedora repositories. To overcome this, I searched the dnf</code> repositories via dnf search nvidia-container-toolkit</code> and found a package named golang-github-nvidia-container-toolkit</code> from the Fedora Project, which pointed to the official NVIDIA GitHub repository for the advertised nvidia-container-toolkit</code> package. I decided that this is a sufficiently trustworthy source for my use case, however you can also install the package from the NVIDIA repository directly, but this will require you adding a new repository to dnf</code>. In my case I ran: sudo dnf install golang-github-nvidia-container-toolkit </code></pre> Generate CDI Specification File</a></h2> With the NVIDIA Container Toolkit we have to generate a CDI (Container Device Interface) Specification File according to the Ramalama installation guide. To do this, simply run: sudo nvidia-ctk cdi generate --output=/etc/cdi/nvidia.yaml </code></pre> You can check if everything worked correctly by running: nvidia-ctk cdi list </code></pre> This should return a list of CDI devices. According to the Ramalama documentation, the CDI Specification file needs to be updated each time the driver is updated or if the GPU configuration changes, so be aware of breaking changes. </blockquote> Serving a Model With Ramalama</a></h2> Once Ramalama is installed and configured, pulling and serving models is relatively simple. For example, to pull the open mistral 7B model, simply run: ramalama pull mistral </code></pre> This will pull the default mistral model specified in the Ramalama configurations. You can however pull specific models from a wide range of model registries, including Huggingface and Ollama. To serve the pulled model, simply run: ramalama --runtime llama.cpp serve -p 8080 mistral </code></pre> This runs the mistral model using the llama.cpp</code> inference runtime (Ramalama also supports vllm</code>) and makes the model accessible locally on port 8080. For this article we use llama.cpp</code>. Before moving on, make note of a line that reads something along the lines of: main: server is listening on http://0.0.0.0:8080</code>, for example, when serving the model. Make note of the http</code> address, this is where the model is accessible locally. Making the Model Accessible in Zed</a></h2> In the Zed IDE, open the settings.json</code> file by opening the top ribbon menu in the top left window, click on Zed</code>->Open Settings</code> and then Edit in settings.json</code>. In settings.json</code> add the following "language_models"</code> key-value pair: { "telemetry":{...}, <- Placeholder field "auto_update":{...}, <- Other placeholder field "language_models": { "openai_compatible": { "Ramalama": { <- Custom name "api_url": "http://0.0.0.0:8080/v1", <- this is where you serve your model "available_models": [ { "name": "Mistral-7B-Instruct-v0.3", <- Custom name "display_name": "Mistral-7B-Instruct-v0.3", <- Custom name "max_tokens": 32768 , <- It will be explained how to get this number }, ], }, }, }, } </code></pre> "api_url"</code>: make sure that this is the same as what is printed to the terminal after serving the model, with a final /v1</code>.</li> </ul> </div> The final /v1</code> endpoint may be different if you decided to use vllm</code> instead of llama.cpp</code>. Make sure to check the documentation of the service that you decide to use. </div> </div> </div> "max_tokens"</code>: To get the "max_tokens"</code> run the ramalama inspect</code> command and look for a line with the word context_length</code>:</li> </ul> ramalama inspect --all mistral | grep context_length </code></pre> This command will tell you the maximum amount of tokens your model (in this case mistral) can accept as context. API Key</a></h2> At this point, if you try to generate text with Zed, the editor may complain that no API key is configured. If this happens, go to the model selection tab (next to the send chat button in the agent panel), select Configure</code> and type none</code> for the Ramalama API key and save. Closing Thoughts</a></h1> If you did all this, Zed should now be able to generate text in-line or using the chat window via Ramalama. You can also chat with GPU support, which may even be faster than using an external paid service. In addition, you get to keep control of your data. I hope you enjoy using Ramalama and that this tutorial was helpful to you! @misc{2026-vsteinborn-ramalama, author = {Victor Steinborn}, title = {How to Use Ramalama to run Containerized LLMs Privately in Zed}, year = {2026}, url = {https://vsteinborn.github.io/posts/ramalama/} } </code></pre> Your Personal Library - How to Use Multiple Wikis With VimWiki in NeoVim 2025-12-15T00:00:00+00:00 Wisdom is not a product of schooling but of the lifelong attempt to acquire it. -- Albert Einstein </blockquote> Motivation</a></h1> Every now and then in life we learn some piece of information we would like to remember. For example, a friend might mention a fun website they recently learned about or we came across an interesting paper we would like to remember for a thesis. It could be great to have some sort of wiki that keeps track of this useful information for us so it is easier to find later. This is exactly what VimWiki</a> attempts to do! With VimWiki you can make your own wiki on topics that interest you. Even more importantly: It's free</li> There is no vendor lock-in (everything can be a markdown</a> file, which is a plain-text file)</li> It's open source (people can audit the code)</li> Your data remains local (important for data privacy)</li> You can interact with your wiki using vim</a> controls (fast, but requires basic vim knowledge)</li> You can have multiple wikis in different files on your computer and have access between them</li> </ul> The last point is exciting because you can have a personal private wiki you keep on your personal computer, and a public 'wiki' about topics that interest you, such as a blog or a way for editing documentation for work. </div> This blog was written with VimWiki! </div> </div> </div> Prerequisites</a></h1> You need to have NeoVim</a> installed on your computer</li> Some basic Vim/NeoVim knowledge is required. It is recommended to run :Tutor<enter></code> after first installing NeoVim and following the tutorial</li> </ul> Installing</a></h1> Note for Vim users: The VimWiki project page has a Vim installation guide. I recommend first installing lazy.nvim</a>, which is a plugin manager for NeoVim. Next, to install the VimWiki plugin I recommend following this guide</a>. This is the best guide I could find and what ultimately helped me install VimWiki. At the end, your ~/.config/nvim/</code> directory should look like this: . ├── init.lua ├── lazy-lock.json <- This file is created when lazy.nvim gets loaded └── lua ├── config │ └── lazy.lua └── plugins └── vimwiki.lua </code></pre> Short Introduction to VimWiki</a></h1> You can start VimWiki within Vim or NeoVim by pressing <leader>ww</code>, where <leader></code> is a key you can set in the configuration files (via vim.g.mapleader</code> in NeoVim). I believe after going through the NeoVim :Tutor</code> tutorial vim.g.mapleader=' '</code>, so by default the <leader></code> key is the space bar in NeoVim. Once in VimWiki you can link pages by writing [[Cool Websites from Friends]]</code>, for example. To follow the link hover over the link in normal mode and press the enter</code> key. This will take you to the Cool Websites from Friends</code> page, which is just another file in your file system. (The extension of the file depends on your settings. It can be a .md</code> markdown file if you like.) You can write text in the file and save the file to create it if it does not exist yet. To go back to the first page, you can do so by pressing the backspace</code> key. </div> You can also use markdown links, such as [Cool Websites](./cool-websites.md)</code>, and follow the links by pressing enter</code> over them in normal mode </div> </div> </div> This is the most basic functionality of VimWiki, but it is quite powerful when coupled with Vim controls. You can quickly travel between sites, edit them and move information between them without touching your mouse. From my experience, this makes for a comfortable and efficient experience when writing notes in meetings for example. Having Multiple Wikis</a></h1> Let's say we want to have two separate Wikis: A personal private wiki</li> A work-related wiki that will be shared with colleagues (e.g. via Dropbox)</li> </ul> We would like the two wikis to be separated and to live in different files. For example, the personal wiki could be in our local \Documents</code> folder and the work wiki can be in a shared Dropbox (which can be accessed as a folder on the computer). To still make use of the speed of VimWiki we can configure VimWiki with two wikis. To do this, modify the vimwiki.lua</code> file: return { "vimwiki/vimwiki", event = "BufEnter *.md", keys = { "<leader>ww", "<leader>wt" }, init = function() vim.g.vimwiki_list = { { -- [[ Personal Wiki ]] name = "Personal Wiki", path = "/path/to/personal/wiki1/", syntax = "markdown", ext = ".md", index = "README" }, { -- [[ Work Wiki ]] name = "Work Wiki", path = "/path/to/work/wiki2/", syntax = "markdown", ext = ".md", index = "SUMMARY" }, } -- The following is needed to force vimwiki use markdown files vim.g.vimwiki_ext2syntax = {['.md'] = 'markdown', ['.markdown'] = 'markdown'} end, } </code></pre> Now when you press <leader>ww</code>, you are taken to the first wiki in vimwiki_list</code>, which is the personal wiki. To access other wikis press <leader>ws</code>, and a menu opens with the available wikis. With the menu open, type 2</code> and then press enter</code> to go to the work wiki. It is also possible to have inter-wiki links by typing [[wn.<Wiki Name>:<Page Name In Wiki>]]</code>. For example, to link to the SUMMARY.md</code> index page in the Work Wiki</code> from the personal wiki, just type [[wn.Work Wiki:SUMMARY]]</code> on any page in the personal wiki. Now you can quickly switch between the personal and work wikis with a few button presses all without touching the mouse! Learning more</a></h1> There are also many other features of VimWiki that make it useful, such as searching for text in a Wiki. If you would like to learn more about VimWiki, I can recommend the well-written documentation, which you can access via the command :h vimwiki</code>. Interestingly, using this setup I can only access the VimWiki documentation after I started VimWiki in NeoVim, so keep this in mind if you are having issues. I hope this guide was helpful to you. Enjoy using VimWiki! If you decide to cite this post, please include the following information. @misc{2025-vsteinborn, author = {Victor Steinborn}, title = {Your Personal Library - How to Use Multiple Wikis With VimWiki in NeoVim}, year = {2025}, url = {https://vsteinborn.github.io/posts/personal-wiki/} } </code></pre> C-Breakout 2025-10-01T00:00:00+00:00 Introduction to Responsible AI 2025-09-30T00:00:00+00:00 Tiny-Mnist-Denoiser 2025-06-01T00:00:00+00:00 Stress Propagation Analysis 2025-03-01T00:00:00+00:00 Top Responsible AI Papers @ Daiki 2024-12-01T00:00:00+00:00 Boids Simulation 2024-12-01T00:00:00+00:00 Docker Jekyll Dev. Environment 2024-07-01T00:00:00+00:00 Rock Paper Scissors 2023-11-18T00:00:00+00:00 Simple Dictionaries 2023-09-01T00:00:00+00:00 Politeness Stereotypes and Attack Vectors: Gender Stereotypes in Japanese and Korean Language Models 2023-06-01T00:00:00+00:00 Open Source Contribution - Added Japanese and Korean Support to Python Library 2023-02-09T00:00:00+00:00 An Information-Theoretic Approach and Dataset for Probing Gender Stereotypes in Multilingual Masked Language Models 2022-06-01T00:00:00+00:00 Jojo GPT2 Bot 2022-03-01T00:00:00+00:00 Solar System Simulation 2017-03-27T00:00:00+00:00

Victor Steinborn

How to Setup Your Personal Home Server - For Scientists and Engineers

How to Use Ramalama to run Containerized LLMs Privately in Zed

Advantages</a></h2> Ramalama and open models are free and open-source</li> You keep your data local and private</li> Open-source models and software can be modified and audited</li>

Things to keep in mind</a></h2> Some models are quite large and require a lot of memory</li>

Your Personal Library - How to Use Multiple Wikis With VimWiki in NeoVim

Having Multiple Wikis</a></h1> Let's say we want to have two separate Wikis:</p> A personal private wiki</li>

C-Breakout

Introduction to Responsible AI

Tiny-Mnist-Denoiser

Stress Propagation Analysis

Top Responsible AI Papers @ Daiki

Boids Simulation

Docker Jekyll Dev. Environment

Rock Paper Scissors

Simple Dictionaries

Politeness Stereotypes and Attack Vectors: Gender Stereotypes in Japanese and Korean Language Models

Open Source Contribution - Added Japanese and Korean Support to Python Library

An Information-Theoretic Approach and Dataset for Probing Gender Stereotypes in Multilingual Masked Language Models

Jojo GPT2 Bot

Solar System Simulation