Mirko BonomiMirko Bonomi

Software Engineer

My AI Homelab

I've been experimenting with running AI workloads locally, building a small homelab around local LLMs, AI-assisted development, automated news research, and generative media.

The goal is to explore what's possible with locally hosted AI models, understand the different tools and workflows available, and gradually build an environment that I can use in my everyday work.

Hardware

The setup currently runs on a Linux laptop equipped with an NVIDIA RTX 5090 GPU with 24 GB of VRAM.

Having a relatively powerful GPU makes it possible to experiment with different local models, image generation, and video generation without relying entirely on cloud-based services.

I'm still experimenting with the setup and continuously exploring different models, configurations, and use cases.

Local LLMs

One of the main components of the homelab is running LLMs locally.

I use llama.cpp as the inference engine and llama-swap to manage and switch between different models depending on the task.

Rather than relying on a single model for everything, I experiment with different models for different purposes.

Some of the models and workflows I'm currently using:

  • Qwen 3.8 27B: Primarily used for coding assistance and general technical tasks.

  • Gemma 4: Used for summarising articles, processing video content, and extracting useful information from longer pieces of content.

The idea is to have different models available for different workloads, without needing to keep all of them loaded in memory simultaneously.

Automated News Research

One of the more interesting experiments is an automated daily news briefing.

I have a Python script that runs once a day, using Gemma 4 to process and summarize information about important developments across three areas:

  • World economics and financial developments.

  • Geopolitics and international affairs.

  • Technology, particularly developments in AI and the broader tech industry.

The goal is to get a concise overview of what's happening in the world without having to manually go through dozens of articles and news sources every day.

It's also a practical way to experiment with LLM-powered information processing and summarisation.

Generative Media

I'm also experimenting with generative media using ComfyUI.

This gives me a visual workflow for experimenting with different image and video generation models, like FLUX 3 Image Klein and Wan 2.2, pipelines, and techniques.

It's an interesting complement to the LLM side of the homelab, allowing me to explore creative AI workloads locally rather than relying exclusively on hosted services.

Development Environment

Most of the development and experimentation happens in Python.

For editing and working on projects, I use Zed as my primary IDE.

Python is also the glue that connects many of the different components, from automation scripts to model integrations and custom workflows.

What's Next?

This homelab is an ongoing experiment. I'm continuously testing new models, exploring different inference configurations, and looking for practical ways to integrate local AI into my everyday workflows.

Some areas I'd like to explore further include:

  • More advanced agentic workflows using local models.

  • Better automation and orchestration between different models.

  • Improving the daily news briefing with better source selection and fact-checking.

  • More experimentation with local image and video generation.

  • Finding additional ways to use local LLMs for software development and research.

I'll keep updating this page as the setup evolves and as I discover new tools, models, and workflows worth documenting.