Careers and industry
What the ML job titles mean, which skills each needs, how interviews and portfolios work, the market at the time of writing, and the regulation an engineer should know about, without legal advice.
Overview
Machine learning work is split across half a dozen job titles whose meanings differ between companies, and the skills interviewed for change every few years. This hub is for engineers planning a move into the field, or from one part of it to another, and for anyone who needs the shape of the regulatory landscape without reading the statutes. Nothing here is legal advice, salary advice or a prediction. Where the page says what the market looks like, it says when it looked.
Key ideas
Roles and what they actually do
The titles cluster into a few jobs. A data scientist answers questions with data: experiments, metrics, causal analysis, dashboards, with modelling as one tool among several. A machine learning engineer builds and ships models and the pipelines around them, and spends most of the week on data, evaluation and serving rather than on architectures. An applied scientist or research engineer sits between research and product, reading papers and reproducing them in the company’s setting. A research scientist publishes, and the job usually requires a doctorate or an equivalent record. An MLOps or ML platform engineer owns the training and serving infrastructure, registries, monitoring and cost. The newest title, AI engineer, means someone who builds applications on top of hosted or open models: prompts, retrieval, agents, evaluation, with fine-tuning as an occasional step. Data engineers build the pipelines everyone else depends on and are the hardest role to fill well.
The common core is the same for all of them: Python, SQL, enough statistics to know when a result is noise, the evaluation discipline in /topics/evaluation-and-benchmarks/, and the ability to write a clear page explaining what was done and why.
Interviews and portfolios
Interview loops for engineering roles at the time of writing usually include a coding round, a machine learning fundamentals round (bias and variance, regularisation, how gradient descent and attention work, how you would evaluate a model for a given task), a system design round (design a recommendation system, a fraud detector, a retrieval-augmented assistant) and a behavioural round.
A portfolio works when it shows judgement rather than volume. Two or three projects that run from a README, with a stated problem, an honest evaluation including what failed, and a short write-up, beat a dozen notebooks that train a model on a Kaggle set. The projects on this site (/projects/tabular-classifier-fastapi/, /projects/tiny-rag-local-embeddings/, /projects/lora-fine-tune-apple-silicon/) are the right size, and the next step is to change the data and the question.
The market, with a date on it
Read on 13 November 2026, the picture was uneven. Demand for engineers who can build and evaluate language-model applications and for people who can run training and inference infrastructure was strong, and compensation at the companies training frontier models was far above the rest of the market. Demand for generalist data scientists had softened from its 2021 peak, and entry-level roles were scarce relative to the number of graduates and career changers. None of this is a forecast.
Regulation an engineer should know exists
Three bodies of regulation shape engineering decisions at the time of writing, and all three change. The EU AI Act (Regulation (EU) 2024/1689) entered into force on 1 August 2024 and applies in stages: prohibitions from February 2025, obligations for general-purpose model providers from August 2025, and most high-risk obligations from August 2026, with some product-embedded systems following in 2027. It classifies systems by their use, not by their technology, so the same model is high risk in a hiring tool and minimal risk in a game. In late 2025 the European Commission proposed delaying parts of the high-risk timetable, and the outcome of that proposal should be checked on the official page rather than assumed from this one.
The second body is export controls. The United States Bureau of Industry and Security has, since October 2022, restricted the export of advanced computing chips and certain model weights to a list of countries, with thresholds defined in terms of performance and interconnect bandwidth. The rules have changed several times since, including a framework announced in January 2025 and rescinded in May 2025. The engineering consequence is that where a training run happens and who may access the weights can be legal questions, and a compliance team should be asked before assuming they are not.
The third is the voluntary frameworks that contracts and auditors reference: the NIST AI Risk Management Framework in the United States, ISO/IEC 42001 for AI management systems, and sector rules that applied to models before any AI-specific law. The UK, at the time of writing, had chosen to regulate through existing sector regulators rather than a single act. None of this requires an engineer to be a lawyer, but it does require keeping the records (training data provenance, evaluation results, model versions, see /topics/mlops-and-deployment/) that every one of these frameworks asks for, and the records are good engineering regardless.
Where to start
Start at the beginning: /learn/beginner/01-what-is-machine-learning/ frames the field, /learn/beginner/06-evaluation-that-does-not-lie/ is the lesson most interview questions are secretly about, and /learn/beginner/10-from-notebook-to-service/ is the skill that separates engineers from notebook users. Then follow the hub for the role you want: /topics/ml-foundations/ and /topics/classical-ml/ for data science, /topics/mlops-and-deployment/ for engineering, /topics/prompting-and-agents/ and /topics/rag-and-retrieval/ for AI engineering, /topics/transformers-and-llms/ and /topics/deep-learning/ for research.
Tools and platforms
| Tool | What for | Licence | Platform |
|---|---|---|---|
| Git and GitHub | Portfolio, contributions and the interview take-home | Various, free tier | n/a |
| Jupyter and VS Code | Exploration and the write-up | BSD-3, MIT | /platforms/opensource/ |
| Hugging Face Hub | Publishing models, datasets and demo Spaces from a portfolio project | Apache-2.0 (tooling) | /platforms/huggingface/ |
| Kaggle | Datasets and competitions, useful for practice and weak as a portfolio on its own | Terms of service | /platforms/google/ |
Common mistakes
- Optimising for the framework of the year instead of fundamentals. Frameworks change every two years. Linear algebra, statistics and evaluation do not.
- A portfolio of tutorial reproductions with no evaluation and no write-up.
- Describing an open-weights model as open source in an interview or a product description. See /topics/open-source-ml/.
- Assuming regulation does not apply because the model is small or the company is. The EU AI Act and sector rules classify by use case.
- Treating a market snapshot, including this one, as a forecast.
Going deeper
- EU Artificial Intelligence Act, official text and timeline, European Commission.
- NIST AI Risk Management Framework, including the generative AI profile.
- Bureau of Industry and Security, advanced computing and semiconductor export controls, United States Department of Commerce.
- Stanford AI Index Report, the annual survey of research, industry and policy data with sources.
- Sculley et al., Hidden Technical Debt in Machine Learning Systems, NeurIPS 2015, still the best description of what the engineering job is.