Quick intro: I'm Samuel. I'm currently writing my master's thesis at BMW and finishing my degree at Hochschule Aalen (Machine Learning & Data Analytics). Both wrap up in the next few months, and at the same time I'm actively looking for my next role inside the company and also outside.

During my time at BMW so far, I've already worked with LLMs in practice. My last project was a Streamlit PoC built on the internal BMW LLM API, deployed on AWS and integrated into Palantir. That experience made two things clear to me. First, the topic isn't hype here, it's productive daily work. Second, my understanding is good enough for PoCs, but for the roles I'm currently eyeing on the job board, I lack depth in the underlying topics. Fine-tuning, model compression, distributed training, MLOps. Those are the skills sitting in the job descriptions I read every day.

Why now

The job market for ML engineers is pretty strong right now. Practically every larger company I know is looking for people who don't just glue APIs together, but actually understand what's happening under the hood. I have the fundamentals, I have some hands-on experience, but between "I've heard of it" and "I can ship it" there's still a real gap on several topics. I want to close that gap deliberately, before the next round of interviews comes around.

Right now would be the worst time to drift into tutorial hell and get lost in theory. I need twelve weeks of focused, practical work. Nothing more.

Why in public

Because learning privately is slower. If I publish a post every week and push a GitHub commit, I have to deliver. Two things I'm hoping for. First, feedback from people who are further along than I am. Second, visibility for the hiring managers who might have my application in front of them in the coming months.

The idea isn't mine. I was inspired by tinystruggles.com, where someone did exactly this kind of public journey in robotics. Different domain, same principle. It worked there. Let's see if it works here.

What's going to happen over the twelve weeks

Roughly three monthly blocks.

Month 1, foundations. PyTorch from the ground up, transformer architecture line by line, the Hugging Face ecosystem. By the end of the month, a nano-GPT I built myself and a benchmark for German-language LLMs.

Month 2, fine-tuning and compression. LoRA, QLoRA, quantization. Actually fine-tuning a model and pushing it to the Hugging Face Hub. Documenting quantization benchmarks honestly.

Month 3, production. MLOps, vLLM, distributed training with DeepSpeed, cloud deployment. A capstone project at the end that brings everything together into something BMW-relevant.

A detailed plan goes at the top of each weekly post.

What to expect from this blog

Every Sunday, a post with three parts. What I learned (concepts, not just tutorials retold). What I built (link to the repo, short explanation of the trade-offs I made). And what didn't work (honestly, because that's often the most valuable part).

No hype, no "10 things you must know about LLMs", no ChatGPT screenshots. Just what actually happened in practice and what I took away from it.

Setup status

This week is week zero. My Linux setup (Ubuntu 24.04, RTX 3060 Ti) is running, PyTorch is talking to the GPU, the repo is live. Tomorrow, week 1 starts with Andrej Karpathy's "Neural Networks: Zero to Hero" alongside learnpytorch.io.

I'm honestly nervous about whether I can keep the pace alongside the thesis and the day job. But if I don't do this now, I never will. If you're working in the same field, subscribe and follow along. Feedback and sparring are very welcome.

See you next Sunday.