[ML Engineer Stories: Getting There] What helped me become an ML engineer

Introduction

Hi, I’m kittchy.

This is for people doing ML in a university lab who are starting to think about jobs. I’ve worked as an ML engineer since April 2024, so two and a half years now.

In this series I’ll write about what I’m glad I did before becoming an ML engineer and in the two and a half years since. That said, it’s only been two and a half years, so I can’t talk big about the industry or how ML engineers should be. What I can write is what I struggled with at the time and what I chose. I write explanations of the techniques themselves on Zenn, so they won’t appear here.

This is the first post, about the things I did while aiming for the job that still pay off today. From the next post on, I’ll move on to what happened after I started working.

What worried me wasn’t the research

As a student I loved research. My topic was domain adaptation for speech recognition, and I presented at domestic meetings and an international conference. I read lots of papers, built things at a hackathon, and interned at three or four companies, where I worked on speech synthesis and MLOps outside my research topic.

Even so, what worried me most wasn’t machine learning but backend, server-side, and infrastructure work. Job titles like “AI engineer” were just starting to multiply. Studying my research field deeply was a given, and I thought I’d struggle unless I could also hold my own as an engineer. In the lab, you’re done once training runs and you get a score. At a company, a model only creates value if it keeps running somewhere. I had never once built that “keeps running somewhere” part.

So between research tasks, I decided to get hands-on with things outside ML. Through internships and part-time jobs I found places outside the lab to write code, and went to see what comes before and after the model. Here’s how that choice turned out.

What still pays off

1. Spreading out sideways, beyond ML

What I learned after joining a company is that engineering makes up 80 percent of an ML engineer’s job (this probably depends on the company). I spend far more time building the place a model runs than training the model. Fetching data, making inference callable, keeping it running, and noticing when it breaks all come with it.

Experience helped more than knowledge. Having done deployment myself, set up CI/CD, and touched Kubernetes became my direct entry point into the work. I had also played with both GCP and AWS, so whichever cloud comes up, I have a map in my head. If that had been blank, I think I’d have stalled every time on the setup before we even got to the machine learning discussion.

My biggest worry was that the company’s server side was Ruby on Rails, and I had never written Ruby. But even across languages, the structure from receiving a request to touching the database is the same, so it was easier than I feared. You can look up how to write things, and the only part I didn’t know was the syntax; I could still read what the code was doing. I think I joined code reviews early not because I could write Ruby, but because I knew what surrounds it.

2. Spending time on papers, not certifications

I earned almost no certifications as a student. Options like the JDLA Deep Learning for GENERAL certificate existed, but I thought what people would look at was whether I’d published papers, so I put my time there.

The same goes for reading. The papers I read beyond what my research needed now give me speed when I follow new techniques. Reading papers no longer takes much energy, so when an unfamiliar technology comes up at work, going to the primary source doesn’t feel like a chore.

What didn’t pay off

If I had to name one, it’s the Fundamental Information Technology Engineer Examination. I don’t regret passing it, but in two and a half years I haven’t yet had a moment where I thought “glad I got that.” What I memorized for the exam only comes up in practice in ways where looking it up when needed is enough.

Closing

What worried me most as a student is, so far, what helps me most. I’m not saying to go light on research. I’m saying that if you see your research through and set aside a little time to work on what’s outside it, life after joining a company gets easier.

If I could say one thing to my past self, it would be: “Take more risks. If you’d tried starting a company or a business, you might have seen a whole new world.” That’s the one thing I never did.

Next time I’ll write about where what I did in the lab paid off at work. See you.