One of my personal goals is to have more actively running services on servers. I have been learning new technologies and at the same exploring different infrastructure options: local homelab, VPS, serverless, renting hardware remotely. Here’s what I’ve explored in Q1-Q2 ‘26:
Homelab Experiments:
I set up a home server lab with a Lenovo Thinkstation with 24GB RAM, Intel i5, 256GB SSD, 4GB VRAM - Nvidia Quadro.
I use Proxmox. It’s cool - that I have connected only 2 cables to the computer - power and ethernet - and i connect using browser from my laptop, using ssh.
Ideas that i explored: 10 containers, some of them:
- Openclaw. Interesting idea, but for my needs - i wanted it to create simple agentic research workflow - it failed and burned too many openai tokens. Wouldnt give access to my email, which seems most hype where is.
- Ollama - installed locally with mainly phi-4-mini
- Local home file cloud - also from browser, accessible from any local computer. Automation on top: using local phi-4-mini model - I can upload any new file to source folder, and then model decides based on file name and file type to which folder to move it.
- Langchain: using mixture of models: Openai API, local models on ollama - research workflows on any topic of interest
- OpenWeb UI - local space for chatting - both openai and local models for chat. Knowledge source - RAG from local cloud integration. Aim: add bielik 11b model. Accept slower working: 4GB VRAM + 7GB RAM model split.
Runpod experiments
Like the idea of renting by hours. 24GB vram 4090 - perfect for many models - I use mostly text - lama, bielik, uncensored models on top of ollama - very easy and fast setup.
- ComfyUI - experimented with it. Loved idea of visualized workflows, wanted to try with text, but seems video is the main focus of this
- I experimented with Terraform - to rent+set up all with one command - it worked partially - the pod setup was failing, so I dropped it.
- Interesting idea to consider: try serverless pods - to have a fav big model activated after sending a request (need to accept high latency, personal use?)
- Conversational agent on top of Bielik - with browser UI - speech to text -> agent -> text to speech. Quite nice - maybe I will use this and explore in future.
VPS
- OVH: Analytical layer on top of Graphmotivo platform: custom-made GTM integration (extended data than simple GTM tags for analytical purposes) -> server receiver -> Bigquery. Also regular Google analytics data for comparison. I love the idea of custom enhanced measurement on user-level, it helps with bot and fraud detection - legitimate purpose.
- GCP VM: graph loading scripts (Bigquery->Neo4j), fraud detection platform frontend (Hackathon teamwork). I started using Claude Code there big time.
How I use Claude Code
I’ve been using Cursor since late 2024, spent a couple of billions tokens since then and had many ups and downs. I tried to learn latest best-practices and tips regularly - they changed big time. Models got so much better.
In January 2026 I tried Github copilot in VScode, but maybe it’s only me, or I had done something wrong with that - but it felt like a transition from a Lamborghini to a tractor.
In May 2026 I installed Claude. I wasn’t convinced of working with AI directly in the terminal, but I’ve got used to it eventually. Got a chance to use it extensively during a Hackathon, and feels like it’s also great, sometimes better than Cursor. It’s a lot clearer to me.
I loved the idea of LLM Wiki - by Andrei Karpathy. That there is a personal wiki that we can build for any topic, and keep documents structured in folders like : /wiki - /entities /ideas /analysis /data /concepts /milestones - custom for every project. User uploads raw documents to /raw folder, then Claude acts like a librarian that reads it, processes it into the wiki. It sticked with me, and there is a lot of space to add more skills, MCPs and optimize for less token usage.
Top comments (0)