Read enough deep learning job ads from Berlin, Amsterdam, Warsaw and Stockholm and the variation is smaller than you would expect. The same handful of frameworks appears, the same deployment tools follow, and the same argument about research versus engineering shows up in the requirements. What deserves your attention is where the ads differ, because those differences reveal which skills hiring managers are actually struggling to find.

This is a reading of the European market as it looks to candidates right now: what postings ask for, how they ask for it, and how to tell a role that will grow your skills from one that will quietly stall them.

The Core Stack Has Mostly Settled

Python remains the entry ticket; nobody debates it. Beyond that, PyTorch is the default in most postings that name a framework at all. TensorFlow still appears, usually in teams whose models have been in production for years and will not be rewritten for the sake of fashion. JAX shows up in research-heavy groups and a few large-scale training teams. Keras survives mainly as a prototyping layer.

What matters more than the logo on your CV is what reviewers assume "PyTorch experience" means. In practice, they look for evidence that you have done the unglamorous parts:

• Written a custom Dataset and dataloader for messy, real data
• Run mixed-precision or distributed training and understood where the time went
• Profiled a slow pipeline and found the bottleneck was I/O, not the model
• Reproduced a paper's result, then said honestly where it fell short
• Exported a trained model to ONNX or TorchScript and served it

A candidate who can talk through two of those items beats one who lists six frameworks.

Deployment Skills Are the Real Filter

Most European postings labelled "deep learning" or "machine learning engineer" are not research jobs. They sit inside product teams, and the requirements reflect that. Docker and Kubernetes appear constantly, sometimes as must-haves. Cloud experience is usually specific: SageMaker, Vertex AI or Azure Machine Learning, plus object storage and a queue somewhere in the flow. Serving layers show up as Triton, TorchServe, Ray Serve, or a home-grown FastAPI wrapper. MLflow, Weights & Biases and DVC cover experiment tracking and versioning; Airflow or Spark covers the upstream data.

The gap employers describe most often is not modelling. It is the ability to take something that works in a notebook and make it work at three in the morning with nobody watching. That means sensible logging, drift and latency monitoring, a rollback plan, and tests that cover data as well as code. If you can show a project with a container, an endpoint, a health check and a metrics dashboard, you are already ahead of most applicants with identical model accuracy.

Skills worth naming explicitly on a CV

Containerising a training job and running it reproducibly on a cluster. Serving a model behind an HTTP or gRPC endpoint with both batch and streaming paths. Quantising or distilling a model to fit a latency and memory budget. Tracking experiments so a colleague can reproduce a run six months later. These are the phrases that survive a ten-second skim.

Research Backgrounds: Valued, But Narrowly

Some fields still weight academic credentials heavily. Medical imaging, robotics, speech, computational biology and foundation-model teams often ask for a PhD or a strong publication record, and they mean it. Elsewhere the framing has shifted towards a specific, testable skill: can you read a recent paper, judge whether it applies to our problem, and implement it without burning a month?

A master's degree with shipped work frequently outperforms a doctorate with only notebooks. Coming from academia, translate your work into engineering terms — dataset size and construction, compute budget, evaluation protocol, what you tried that failed. Coming from industry, show you engage with the literature rather than only with a framework's API. Fine-tuning, evaluation design and dataset curation now carry as much weight as architecture choices.

Where Demand Concentrates

Geography shapes the stack more than most candidates expect.

• Germany — industrial and automotive work plus a busy Berlin startup scene; some roles still expect German, though English-only teams are common in larger cities.
• Netherlands and the Nordics — product companies and platform teams, usually English-language workplaces, strong on infrastructure and operations.
• France and Switzerland — research-intensive labs and robotics, with more emphasis on mathematics and publications.
• UK and Ireland — finance, healthcare and European offices of large technology firms; hiring cycles here tend to move faster.
• Poland, Spain and Portugal — growing engineering hubs where the work is often production-focused rather than research-focused.

Visa and immigration rules differ by country and change often. Check official government sources or a qualified adviser rather than trusting forum posts from three years ago.

How to Read a Posting in Ten Minutes

1. Sort requirements into must-haves and wishlist. Adverts rarely label them; the ordering usually gives it away.
2. Count the verbs. "Train", "serve", "monitor" and "own" signal a product role. "Publish", "investigate" and "novel" signal research.
3. Look for evaluation language. Teams that mention offline metrics, controlled tests or error analysis usually know what they are doing.
4. Check what the team owns. If that is unclear in the ad, it will be unclear in the job.
5. Find your honest gap. If you match most of the list but not the deployment half, that is your next project.

A Practical Plan for the Next Quarter

Pick one framework and go deeper rather than wider. Take a dataset you find genuinely interesting, train something reasonable, then spend as much time serving it as you did training it: a container, an endpoint, logging, a simple dashboard, and a short README explaining the trade-offs you made. Write up two experiments that failed and what you learned from them. That single repository will do more for your applications than a longer list of finished tutorials.

Then apply selectively. Ten well-matched applications, each with a paragraph connecting your project to the team's actual problem, outperform a hundred generic ones. Ask for feedback after rejections, keep a short log of what interviewers probe, and let that log decide what you study next. The stack will keep shifting; the habit of building, shipping and explaining your work will not go out of demand.

Photo: Tumisu / Pixabay