A complete path to becoming a machine learning specialist, the person who builds, trains, deploys, and maintains models that actually work in production.
This roadmap goes far beyond fitting a model in a notebook. It takes you through the full breadth of modern machine learning: classical algorithms, deep learning, natural language processing, computer vision, big data tooling, MLOps, and the large language models reshaping the field today.
What this role is called
Machine Learning Engineer is the standard title, but the emphasis shifts considerably between companies advertising it.
The core titles are Machine Learning Engineer, ML Engineer and Applied Machine Learning Engineer. These usually mean the full loop: building models, getting them into production, and keeping them healthy.
The engineering-led titles skew toward the second half of this roadmap: MLOps Engineer, ML Platform Engineer, ML Infrastructure Engineer, Production ML Engineer. If the MLOps and big data stages appeal most, these are your titles, and they are in unusually short supply.
The specialist titles narrow to one domain: Computer Vision Engineer, NLP Engineer, Recommendation Systems Engineer, Search Engineer. Each takes this roadmap’s foundations and goes deep in one direction.
The progression above is Senior ML Engineer, then Staff or Principal on the individual contributor track, or ML Engineering Manager on the other.
Two neighbours are worth distinguishing. Data Scientist overlaps on modelling but stops short of deployment and monitoring, and is the role this roadmap builds on rather than competes with. LLM Engineer goes deeper into model internals and architecture rather than broader across the production stack, and datalad has a separate roadmap for it.
How long it takes
This roadmap assumes you have already completed the Data Scientist Roadmap, which is where the mathematics, statistics and Python fluency come from. The section below explains why that matters.
With that foundation, expect nine to fifteen months part-time. The core machine learning stage moves quickly because you already have the statistics, and the time goes into deep learning, the specialised domains, and above all MLOps.
Coming from a working data scientist role, expect six to nine months, and expect most of it to be the engineering half. You can already build models; what you are adding is Docker, CI/CD, experiment tracking, monitoring and pipelines.
Coming from software engineering with some machine learning exposure, expect twelve to eighteen months. The engineering stages will be fast and the modelling stages slow, which is the mirror image of the data scientist’s path.
Starting from scratch without the Data Scientist Roadmap is not recommended, and the next section explains why in more detail.
What it pays
These are indicative annual ranges. Machine learning engineering pays above data science in most markets, and the premium comes from the production skills rather than the modelling ones.
United Kingdom. Entry-level roles start around £45,000 to £60,000. Mid-level runs roughly £65,000 to £90,000. Senior engineers sit around £95,000 to £130,000, and staff or lead roles reach £130,000 to £180,000.
Europe. Germany and the Netherlands run mid-level at roughly €70,000 to €95,000 and senior at €100,000 to €140,000. Switzerland sits clear of the rest, and remote roles for US companies distort the top of every European range.
United States. Entry-level runs roughly $110,000 to $140,000. Mid-level sits around $140,000 to $190,000. Senior roles reach $190,000 to $260,000, and staff or principal roles at large technology companies go substantially higher once equity is counted.
Contract and freelance. UK day rates typically run £500 to £1,000, and US rates $800 to $1,500, with the premium going to people who can take a model to production rather than only build it.
The pattern is consistent across every market: the MLOps stage is what moves you up the range. Model building is now widely taught and increasingly assisted by tooling, while deploying, monitoring and versioning models reliably remains scarce.
Who this roadmap is for
It suits data scientists whose models keep dying in notebooks, software engineers moving into machine learning who want the modelling depth to match their engineering, and anyone who has finished the Data Scientist Roadmap and wants to build things that run.
How each lesson works
Every lesson is a main article, plus whichever of these fits it:
Video is the walkthrough. Cheatsheet is a one-page reference to keep beside you. The math is the derivation underneath the method, for the lessons where the mathematics changes how you use it. Tools are interactive generators and explorers you can use while you learn. Projects in the real-world stage also come with the dataset they run on.
Not every lesson has every component. The chips on each card show what is available, and a greyed-out chip is one that is still to come.
Machine learning is applied statistics built on solid engineering, and trying to learn it without that groundwork leaves you running algorithms you cannot reason about. Finish the Data Scientist Roadmap, and you arrive here ready to build on a foundation that holds. Begin here without it, and you will spend half your time backfilling concepts you should already own.
datalad roadmap
Machine Learning Scientist / Engineer
Twelve stages, 111 steps, from the classical algorithms to models running in production.
Finish the Data Scientist Roadmap first
PrerequisiteThis roadmap assumes the mathematics, statistics and Python fluency that the earlier path gives you. Begin here without it and you will spend half your time backfilling concepts you should already own. The glossary is not a prerequisite, it is a reference to keep open while you work.
The mathematics, statistics and Python fluency this roadmap assumes
175 terms from AUC to zero-shot learning, for the words a lesson assumes you know
From core algorithms to production
12 stagesWork through the stages in order. It is tempting to rush toward the deep learning and the LLMs and skip the engineering, but deploying, monitoring and versioning a model is what makes you an engineer rather than someone who fits models in notebooks.
Builds on Multi-modal models on Hugging Face
Builds on the Product Catalogue Auto-Tagging project
Builds on Combining models with ensemble methods
Builds on the Anomaly Detection When Positives Are Rare project
Builds on Getting started with Hugging Face
Builds on Machine learning at scale with PySpark
Builds on the MLE-2 Serve the Model project
Builds on the MLE-3 Batch Scoring Pipeline project
Builds on the MLE-7 Upstream Change project
Builds on Workflow orchestration with Airflow
Builds on the MLE-4 Offline and Online Parity project
Builds on Tracking experiments with MLflow
Builds on the Demand Forecasting With Backtesting project
Builds on Making models explainable
Builds on Fairness and bias in machine learning
The model with the best precision recovers the least money. Ranking by expected value finds 2.7 times more from the same calls
The score multiplies conversion by 1.82. Answering inside six hours multiplies it by 5.36. The model was the smaller half
Clean segments, stable under resampling, and six months later only 37 percent of customers were still in them
A rolling origin backtest, a seasonal naive baseline that is hard to beat, and the metric that ranked the worst line as one of the best
Two years of prices and units, a clean log-log fit, and an elasticity 45 percent too small. Every control gets you closer without getting you there
Six recommenders on one holdout. The best hit rate fills its slots with bestsellers and reorders. The worst is worth eleven times more
Fraud is 0.44 percent of transactions, so flagging nothing is 99.56 percent accurate. Building a review queue that is 5 percent precise anyway
Conversion rose 2.98 points and the campaign lost money. Targeting the likeliest buyers made it worse
Three architectures, ten splits, equal tuning budgets. The simplest wins, and by less than a quarter of the noise between splits
A word counter looks finished at 89.7 percent until the wording changes. The best possible score is 93.1, and knowing that changes the decision
Topic modelling recovered four of the ten real themes and missed the complaint that tripled. An afternoon of labelling beat it
Six attributes, one model, and a 65 point spread hidden inside a single average. One of the six should not ship at all
Four detectors held to one false alarm budget all caught the same eight. Removing the day of week was worth four times the LSTM
Rules reading three quarters of the fields scored zero on a new supplier. Five label strings fixed it, and arithmetic caught the rest
Use this Roadmap anywhere
Free to share and embed, no attribution required, though a credit link is appreciated and included by default. Drop the image straight into Notion, Slack or your own site.
How to Use This Roadmap
Work through the stages in order. The core machine learning stage gives you command of the classical algorithms and the workflow around them: preprocessing, feature engineering, validation, and tuning. From there, deep learning and the specialised domains of NLP, vision, and reinforcement learning expand what you can build. The big data and MLOps stages are what separate a model that works on your laptop from one that runs reliably in production, serving real users and staying healthy over time. And the final stages on large language models and explainable AI bring you to the frontier of the field.
A word on emphasis: it is tempting to rush toward the exciting topics, the LLMs and the deep learning, and skip the engineering. Resist that. The ability to deploy, monitor, version, and validate a model is what makes you an engineer rather than someone who fits models in notebooks. Those skills are also where much of the real demand sits. Give the MLOps stage the attention it deserves, and you will be the rare specialist who can take a model all the way from idea to production and keep it running.
Happy learning.
[…] Soft voting requires every model to produce probabilities, which means a small adjustment for support vector machines. […]