Machine Learning Scientist / Engineer Roadmap

This roadmap details the comprehensive journey to becoming a machine learning specialist, emphasising foundational skills, core algorithms, deep learning, MLOps, and the significance of responsible AI in production environments.

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:

Code along is a hands-on set of examples you run yourself. Cheatsheet is a one-page reference to keep beside you. Feynman Technique is the same idea explained as simply as it can honestly be put. The math is the derivation underneath the method, for the lessons where the mathematics changes how you use it. Video is the walkthrough.

Not every lesson has every component, and the list under each one tells you what is available.

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

Nine stages, 69 steps, from the classical algorithms to models running in production.

9 to 15 months with the foundations done6 to 9 from a working data scientist12 to 18 from software engineering

0 of 69 steps complete

Start here

Finish the Data Scientist Roadmap first

Prerequisite

This 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.

Curriculum

From core algorithms to production

9 stages

Work 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.

01

Core machine learning

The classical algorithms and the workflow around them

0 of 13 steps done
Decision trees and tree-based models
ArticleCode alongCheatsheetFeynmanThe mathVideo
Clustering and cluster analysis
ArticleCode alongCheatsheetFeynmanThe mathVideo
Machine learning on time series
ArticleCode alongCheatsheetFeynmanThe mathVideo
Combining models with ensemble methods
ArticleCode alongCheatsheetFeynmanThe mathVideo
02

Natural language processing

Turning language into something a model can use

0 of 3 steps done
Working with text: NLP foundations
ArticleCode alongCheatsheetFeynmanVideo
Practical NLP using spaCy
ArticleCode alongCheatsheetVideo
Turning text into features
Embedding explorerArticleCode alongCheatsheetFeynmanThe mathVideo
03

Deep learning

PyTorch and Keras, from first tensor to transformer

0 of 11 steps done
Working with images: image processing
ArticleCode alongCheatsheetVideo
Image models in PyTorch
ArticleCode alongCheatsheetFeynmanThe mathVideo
Text models in PyTorch
ArticleCode alongCheatsheetFeynmanThe mathVideo
Building image models in Keras
ArticleCode alongCheatsheetVideo
Sequence models and RNNs in Keras
ArticleCode alongCheatsheetFeynmanVideo
Deep learning cheatsheet, all 23 topics on one page
04

Reinforcement learning

Learning from reward rather than from labels

0 of 3 steps done
Reinforcement learning with Gymnasium
ArticleCode alongCheatsheetFeynmanThe mathVideo
Deep reinforcement learning
ArticleCode alongCheatsheetFeynmanThe mathVideo
Reinforcement learning from human feedback (RLHF)
ArticleCode alongFeynmanThe mathVideo
05

Big data and distributed computing

When the data stops fitting on one machine

0 of 6 steps done
Getting started with PySpark
ArticleCode alongCheatsheetFeynmanVideo
Machine learning at scale with PySpark
ArticleCode alongCheatsheetVideo
Getting started with Databricks
ArticleCode alongCheatsheetVideo
Querying data with Databricks SQL
ArticleCode alongCheatsheetVideo
Managing data in Databricks
ArticleCode alongCheatsheetVideo
Visualising data in Databricks
ArticleCode alongCheatsheetVideo
06

MLOps and engineering for ML

The stage that separates a notebook from a product

0 of 8 steps done
Command line essentials for ML
ArticleCode alongCheatsheetVideo
Tracking experiments with MLflow
ArticleCode alongCheatsheetVideo
Building data pipelines: ETL and ELT
ArticleCode alongCheatsheetFeynmanVideo
Validating data quality with Great Expectations
ArticleCode alongCheatsheetVideo
Versioning data with DVC
ArticleCode alongCheatsheetVideo
Monitoring models in production
Drift generatorArticleCode alongCheatsheetFeynmanVideo
Containerising work with Docker
ArticleCode alongCheatsheetFeynmanVideo
CI/CD pipelines for machine learning
ArticleCode alongCheatsheetVideo
07

Large language models and generative AI

The frontier, and the tooling that makes it usable

0 of 6 steps done
Building AI agents with smolagents
ArticleCode alongCheatsheetFeynmanVideo
Optimising Llama 3 for your own tasks
ArticleCode alongCheatsheetFeynmanThe mathVideo
08

Responsible and interpretable AI

Being able to explain and defend what the model did

0 of 2 steps done
Making models explainable
ArticleCode alongCheatsheetFeynmanVideo
Fairness and bias in machine learning
ArticleCode alongFeynmanVideo
09

In the real world

Business problems worked end to end, machine learning first and deep learning after, each with the finding that surprised the room

0 of 15 steps done
Late Payment Prediction

The model with the best precision recovers the least money. Ranking by expected value finds 2.7 times more from the same calls

ArticleDatasetVideo
Lead Scoring and Threshold Economics

The score multiplies conversion by 1.82. Answering inside six hours multiplies it by 5.36. The model was the smaller half

ArticleDatasetVideo
Segmentation That Survives Scrutiny

Clean segments, stable under resampling, and six months later only 37 percent of customers were still in them

ArticleDatasetCheatsheetVideo
Demand Forecasting With Backtesting

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

ArticleDatasetVideo
Price Elasticity From Observational Data

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

ArticleDatasetVideo
Recommendations From Order History

Six recommenders on one holdout. The best hit rate fills its slots with bestsellers and reorders. The worst is worth eleven times more

ArticleDatasetVideo
Anomaly Detection When Positives Are Rare

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

ArticleDatasetCheatsheetVideo
Uplift Modelling From an A/B Test

Conversion rose 2.98 points and the campaign lost money. Targeting the likeliest buyers made it worse

ArticleDatasetCheatsheetVideo
Deep Learning Against Gradient Boosting

Three architectures, ten splits, equal tuning budgets. The simplest wins, and by less than a quarter of the noise between splits

ArticleDatasetVideo
Support Ticket Triage

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

ArticleDatasetCheatsheetVideo
Review Themes at Scale

Topic modelling recovered four of the ten real themes and missed the complaint that tripled. An afternoon of labelling beat it

ArticleDatasetCheatsheetVideo
Product Catalogue Auto-Tagging

Six attributes, one model, and a 65 point spread hidden inside a single average. One of the six should not ship at all

ArticleDatasetVideo
KPI Anomaly Alerting

Four detectors held to one false alarm budget all caught the same eight. Removing the day of week was worth four times the LSTM

ArticleDatasetVideo
Invoice and Receipt Extraction

Rules reading three quarters of the fields scored zero on a new supplier. Five label strings fixed it, and arithmetic caught the rest

ArticleDatasetCheatsheetVideo
Article, the main lessonTool, live now and free to useCode along, examples you run yourselfCheatsheet, one page to keep beside youDataset, the file the project runs onFeynman, the idea in plain EnglishThe math, the derivation underneathGreyed out means not published yet
Share and embed

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.

View Comments (1)

Leave a Reply

Subscribe to My Newsletter

Subscribe to my email newsletter to get the latest posts delivered right to your email. Pure inspiration, zero spam.

Discover more from Discuss Data Science, Machine Learning and Analytics

Subscribe now to keep reading and get access to the full archive.

Continue reading