mrdjan.net
Contact

← Back to Blog

AIcareer breakragfrontendjs fatiguellmopsdeveloper journeyself-reflection

Career Break in Tech: Why I Took One and What I Built

Why I took a career break, what I built with AI and RAG systems, and lessons from ~5 months of experimenting.

•5 min read
Mrdjan StajicSenior Frontend Engineer with 10+ years of experience building scalable web applications. Passionate about accessibility, performance, and modern web technologies.

I took a career break

Career break in 2025 & Age of AI

After budgeting and tracking expenses for a year and a half, I did it

Advice

  • The global job market is not good; too many good candidates
  • Do not do this on a whim
  • I spent a year tracking my budget
  • Worked 10 years with ~4 months in between jobs
  • Good at planning and Excel sheets
  • Think about it

A couple of reasons why

  • I was tired
  • When I started working in 2015, it was normal for people to take career breaks. I knew people from Serbia and other countries who went traveling for 6–9 months around the world.
  • JS Fatigue (that's a term). I'm mainly a frontend dev (JS/TS world), low on backend production projects, but with a lot of side projects.
  • I wanted to explore AI — let's call that a Bet

I was tired

Self-explanatory. I really felt I was stagnating in my career and wanted to learn something else, but the options for branching weren't interesting to me.

What I did while I didn't code or explore

  • I read — mostly Sanderson (Stormlight Archive, Mistborn Era 2 read)
  • Played video games, RDR2 (rerun), Clair Obscur: Expedition 33 — even if you don't play games, check out the soundtrack, it's on Spotify
  • Coded for myself, related to the Bet

JS Fatigue

  • I have been in the industry for 10 years, mostly web, full-stack/frontend
  • In that time period, I've seen the loop: everything on the server, then with JS, then SPA, then again resurgence of server-rendered content   - Main difference: the cloud
  • Only abstractions are added, and the ways of fetching data have changed — the end result is the same
  • In my opinion, there are more abstractions now. What really changed is how we draw, fetch data, and interact with the server
  • For frontend devs, it's now normal to know more than just frameworks: cloud knowledge, CI/CD, etc., a bit of infra.

The Bet

  • There is a big push in AI, heavily subsidized by Microsoft, Tencent, and private VCs. If you don't know that, you're living under a rock
  • ROI must be achieved, so companies are pushing AI-assisted tools — my thought
  • People in C-level positions at prominent companies mention that this is a bubble (duh), but I take their opinions with a grain of salt
  • The LLM world, as of the time of writing (just after ChatGPT 5 went live), has plateaued — my opinion
  • The cycle of bubbles is always there: dot-com, crypto, maybe AI is next
  • Nobody knows what will happen in 5–10 years
  • Chip providers are the choke point for cheap inference
  • If the bubble bursts, my bet is that it'll be cheaper to rent a GPU and deploy your own infra; also, privacy

What I made

  • Blog analyzer Enter a blog post address in a CLI; it uses two LLMs (Terrasect and Qwen3) for OCR and summarization. Python for web scraping Blog Analyzer

  • D&D Agent with TUI, with Fast MCP for rolling dice, attacks, spawning items, and storing local history. Single-player D&D adventure creator that works on Qwen Instruct with tool calling and local Fast MCP for navigating the world with narrative storytelling

  • Obsidian → RAG   I use Obsidian as a brain dump, so I first implemented barebones RAG with cosine similarity, which worked well for structured data. For unstructured data, I hardened it via libraries, and the results are better but not perfect. Obsidian RAG GitHub repo

  • Jira (UI & Barebones MCP)   - Barebones MCP that can be plugged-and-played into VS Code (for example) and works via stdio   - Has a bridge to Nest, which is connected to the UI in a SPA Vite app   - Uses MCP to get the current data via HTTP bridge   - Uses Gemma-Instruct for fuzzy calls to map them to real tool calls   - Tool call returns JSON (task, issues, sprint, board, epic)   - Chat about the risks, list subtasks, etc. via LLM (Qwen2.5)

  • That video on the top ThreeJS scene maker   - Using ThreeJS ChatGPT 5 for initial prompt.   - Opened Codex, added prompt as an MD   - Burned 1.9M tokens in two conversations until I made it   - Two phases: first with voxels   - Models   - Will write a blog post about it

What I have learned

  • How to find a model that actually works
  • Hugging Face is an awesome place
  • Downloading a GGUF is sometimes easier than building it
  • llama.cpp maintainers are awesome
  • Need to explore more stuff like vLLM
  • Quantizing models to run on my hardware
  • Model orchestration
  • Llama hell
  • Different model, different template hell
  • Everybody hallucinates
  • Stuff can be delivered as MVPs very easily

How I did it

  • Using only AI-assisted coding
  • Mostly deleting code, writing prompts, and changing environment variables and infra setup
  • Will go over every example and maybe add some about me in the future
  • I did not code at all (well, 98%)
  • Max amount spent monthly: 40 bucks

Thanks for reading

Best of vibes,

Mrdjan

For contact, you can use LinkedIn or Contact

About the Author

Mrdjan Stajic

Senior Frontend Engineer with 10+ years of experience building scalable web applications. Passionate about accessibility, performance, and modern web technologies.