Data & AI · Updated 16 September 2026
Machine Learning Engineer salary and career roadmap
Build, train and run models in production - the software engineering job where the software happens to learn.
Salary data checked · outlook from BLS projections released · by Bilal Tahir
- What it is
- A Machine Learning Engineer is a software engineer who owns machine learning models in production: the pipeline that assembles training data, the evaluation that decides whether a new model ships, the serving path that answers in milliseconds under load, and the monitoring that catches drift.
- Salary
- A Machine Learning Engineer in the United States earns a median of $140,000 entering the field, $200,000 at mid-career and $270,000 at senior level; the top end is $497,500. (Levels.fyi United States machine learning engineer submissions, widened with Glassdoor and BLS Occupational Employment and Wage Statistics May 2025, )
- Outlook
- Employment is projected to grow 10% over the ten years to 2036, with about 106,100 US openings a year. (BLS Occupational Outlook Handbook, 15-1252 Software Developers (closest official proxy; the BLS publishes no machine learning engineer occupation), )
- Time to first job
- A career changer starting from zero typically needs 36 to 60 months at 10-15 hours a week to reach a first offer.
- Cost to get in
- The cheapest verified route in (Data scientist to ML engineer) costs about $800; the most expensive costs about $9,000.
- Roadmap
- Our Machine Learning Engineer roadmap is 8 steps and about 1,360 study hours; the fastest route in is Data engineer to ML engineer at about 18 months.
- Degree
- No degree is legally required, but Machine Learning Engineer is the least degree-flexible of the data roles in practice: most large-employer postings list a bachelor’s degree in computer science or another quantitative field, and a master’s is still the standard credential for new entrants. A portfolio substitutes for it only when it contains deployed systems or open-source contributions, not notebooks.
- Automation exposure
- Low. Code assistants now write much of the boilerplate a Machine Learning Engineer used to type, along with first-draft pipelines, tests and documentation, which is why the same loops now include an AI-assisted coding round. What does not automate is scoping the problem, designing the evaluation that decides whether a model ships, and owning a system whose failures are statistical rather than a stack trace.
What does a Machine Learning Engineer do?
A Machine Learning Engineer is a software engineer who owns machine learning models in production: the pipeline that assembles training data, the evaluation that decides whether a new model ships, the serving path that answers in milliseconds under load, and the monitoring that catches drift. A machine learning engineer owns models in production. The same work is also posted as ML Engineer, Machine Learning Software Engineer, Applied Scientist, MLOps Engineer.
A machine learning engineer owns models in production. That means the pipeline that assembles training data, the training job itself, the evaluation harness that decides whether a new model is better than the old one, the serving path that answers in 50 milliseconds under load, and the monitoring that catches drift before a stakeholder does. At large companies you will specialise - ranking and recommendations, ads, search relevance, fraud and risk, speech or vision, or the platform team that builds the training infrastructure everyone else uses. At small companies you are the entire ML function, which means you also do the data engineering.
The critical thing to understand about this role is that it is a software engineering job first. The bar in interviews is the same data-structures-and-algorithms bar as a backend engineer, plus ML system design on top. People who come in from research and cannot write clean, tested, deployable code struggle; people who come in from backend engineering and learn the modelling usually do better than the reverse. Chip Huyen's line that most ML failures in production are data and infrastructure failures rather than model failures is the honest summary of the day-to-day.
It pays extremely well - Levels.fyi puts the US median total compensation at $280,000 and the 90th percentile at $497,500 - and it is the hardest of the four data roles to enter. Almost nobody gets hired into it directly from a career change. The realistic paths are software engineer to ML engineer, or data scientist to ML engineer, and both take two to four years. Also be honest about the distribution: that $280,000 median is a Levels.fyi figure dominated by large tech employers, while Glassdoor's broader-market average for the same title is $164,714.
Why Machine Learning Engineer pay is high
Three forces stack. First, the skill combination is genuinely scarce: production-grade distributed systems engineering plus statistical modelling plus enough infrastructure knowledge to reason about GPUs, memory and latency. Most engineers have one of those, few have all three. Second, the leverage is enormous and directly measurable - a recommendation or ads ranking model that improves by half a percent can be worth eight or nine figures a year at a large platform, so the employer's willingness to pay is bounded by revenue rather than by an HR band. Third, since 2023 the largest technology companies and well-funded AI labs have been bidding against each other for the same pool, and they pay in equity: median base on Levels.fyi is $200,000 but median total compensation is $280,000, and at Meta the median machine learning engineer package is $458,000 with the 90th percentile at $849,600. That equity premium is also the risk - a large share of the headline number is stock whose value you do not control.
What's good
- Among the highest-paid individual contributor roles outside of finance: Levels.fyi US median total compensation of $280,000 and 90th percentile of $497,500.
- Work is intellectually dense and the feedback is objective - the model is better or it is not.
- Skills transfer cleanly to AI engineering, data engineering, backend engineering and research infrastructure.
- Strong bargaining position: the skill combination is scarce and demand is broad across industries.
- Direct, measurable business impact, which makes promotion cases straightforward to write.
- Remote roles are common at AI-native companies, often at a single national pay band.
What's hard
- The hardest of the four data roles to enter - realistically two to four years from a related job, not a direct career change.
- You are held to a full software engineering interview bar on top of the ML content.
- On-call for systems that fail statistically, which is far harder to debug than a crash.
- Much of the day is data plumbing and infrastructure, not modelling.
- Headline compensation is equity-heavy and concentrated in a small number of employers, so it is cyclical and partly outside your control.
- Fast-moving field: a meaningful fraction of what you know has a three-year half-life.
What a Machine Learning Engineer does all day
- 09:00 - Check the overnight training run: it OOM'd on the third epoch after a schema change upstream added a column.
- 09:30 - Fix the feature pipeline, add a schema assertion so it fails loudly next time, open the PR.
- 10:30 - Design review for a new ranking model: argue about the offline metric and whether it will correlate with the online one.
- 11:30 - Code review two PRs from teammates; one is a training/serving skew waiting to happen and you say so.
- 13:00 - Deep work: profile the inference service, find that tokenisation is 40% of p99 latency, batch it, halve the number.
- 15:00 - Pair with a data scientist to turn their notebook feature into a tested pipeline component.
- 16:00 - Launch a shadow deployment of the new model and set up the comparison dashboard.
- 17:00 - Write the weekly update: what shipped, the online A/B numbers, what is blocked on GPU capacity.
- Occasionally - On-call: a model starts returning garbage at 2am because an upstream feed changed format. Roll back first, diagnose after.
Machine Learning Engineer salary in 2026: by level
US, annual, USD. Base plus typical bonus where the source reports it.
A US Machine Learning Engineer earns a median of $140,000 entering the field, $200,000 at two to four years and $270,000 at senior level, with the top end at $497,500. These are base-salary figures in US dollars as of September 2026, synthesised from Levels.fyi United States machine learning engineer submissions, widened with Glassdoor and BLS Occupational Employment and Wage Statistics May 2025. Pay varies about 20-40% by metro.
Two datasets tell two different stories and you need both. Levels.fyi US machine learning engineer total compensation (retrieved September 2026): 10th percentile $150,000, 25th $199,999, median $280,000, 75th $382,810, 90th $497,500 - with base at $200,000 median and equity at about $51,875 median, so roughly a third of the median package is stock. Glassdoor's broader sample puts the US average at $164,714 with a 25th percentile of $132,390, 75th of $207,578 and 90th of $254,003; that is closer to what a non-tech employer, a regional bank or a mid-size SaaS company pays. Company medians on Levels.fyi: Meta $458,000 (10th $271,000, 90th $849,600), Apple $386,000, Google $288,000 (L3 from about $199,000 to L7 at about $743,000), Amazon $280,000 (L4 about $177,000 to L6 about $483,000), Nvidia $261,000. BLS has no ML engineer code; the closest are Software Developers (15-1252, $135,980 median May 2025, 10th $82,460, 90th $214,670) and Computer and Information Research Scientists (15-1221, $140,300 median). Geography still matters: Bay Area and Seattle packages run 20-40% above national, though remote ML roles at AI-native companies increasingly pay a single national band. Assume the Levels.fyi figures apply if you are interviewing at a public tech company or a funded AI startup, and the Glassdoor figures if you are not.
BLS does not publish a machine learning engineer occupation. Most ML engineers are classified as Software Developers (15-1252): 1.9 million employed in 2025, +10% projected 2025-2035, about 106,100 annual openings, with BLS naming artificial intelligence as a demand driver. The research-heavy end maps to Computer and Information Research Scientists (15-1221): 38,600 employed, +22% projected, about 2,900 annual openings. Data Scientists (15-2051) is +35% with 24,800 openings. The true growth rate for ML-specific engineering is almost certainly at the top of that range, but no official series isolates it. Automation risk is low for the role and high for parts of the work: code assistants write a great deal of the boilerplate, but deciding what to build, designing the evaluation, and owning a system that fails in statistical rather than deterministic ways is the job. The real risk is not automation but concentration - a large share of the highest-paying roles sit at a small number of employers, and that market is cyclical.
“Demand for software developers, software quality assurance analysts, and testers is projected to be strong due to the continued expansion of software development for artificial intelligence (AI), Internet of Things (IoT), robotics, and other automation applications.”
Machine Learning Engineer salary by city
National bands scaled by metro wage differentials from the BLS May 2025 OEWS release.
Machine Learning Engineer pay is highest in San Jose / Silicon Valley (mid-career median about $284,000, ×1.42 the national figure) and lowest among large metros in Salt Lake City (about $190,000). The multiplier moves the offer, not what you keep after rent and state tax.
| Metro | Entry | Mid | Senior | vs national |
|---|---|---|---|---|
| San Jose / Silicon Valley | $198,800 | $284,000 | $383,400 | ×1.42 |
| San Francisco Bay Area | $189,000 | $270,000 | $364,500 | ×1.35 |
| New York City | $179,200 | $256,000 | $345,600 | ×1.28 |
| Seattle | $168,000 | $240,000 | $324,000 | ×1.2 |
| Boston | $161,000 | $230,000 | $310,500 | ×1.15 |
| Washington DC metro | $156,800 | $224,000 | $302,400 | ×1.12 |
| Los Angeles | $151,200 | $216,000 | $291,600 | ×1.08 |
| Chicago | $147,000 | $210,000 | $283,500 | ×1.05 |
| Austin | $147,000 | $210,000 | $283,500 | ×1.05 |
| San Diego | $145,600 | $208,000 | $280,800 | ×1.04 |
| Denver | $142,800 | $204,000 | $275,400 | ×1.02 |
| Philadelphia | $142,800 | $204,000 | $275,400 | ×1.02 |
| Dallas | $140,000 | $200,000 | $270,000 | ×1.0 |
| Minneapolis | $140,000 | $200,000 | $270,000 | ×1.0 |
| Raleigh-Durham | $140,000 | $200,000 | $270,000 | ×1.0 |
| Houston | $138,600 | $198,000 | $267,300 | ×0.99 |
| Atlanta | $137,200 | $196,000 | $264,600 | ×0.98 |
| Phoenix | $133,000 | $190,000 | $256,500 | ×0.95 |
| Miami | $133,000 | $190,000 | $256,500 | ×0.95 |
| Salt Lake City | $133,000 | $190,000 | $256,500 | ×0.95 |
A multiplier raises the number on the offer letter, not what you keep. San Francisco pays about 35% more than the national median for these roles, but median Bay Area rent and California state income tax eat most of that for anyone below the senior rung. Texas, Florida, Washington, Tennessee and Nevada levy no state income tax, which is worth roughly 4-10% of take-home versus California or New York City, where city tax stacks on top of state tax. Run the comparison on after-tax income minus housing before you move. Fully remote roles are the edge case worth chasing: a national pay band spent in a 0.88 cost market beats a 1.35 salary spent in a 1.6 cost market for most people.
How the multipliers are derived
Anchor: the BLS May 2025 OEWS national mean wage across all occupations is $33.54/hr ($69,770/yr). Metro all-occupation means from the same release: San Jose-Sunnyvale-Santa Clara $57.32 (1.71x national), San Francisco-Oakland-Fremont $48.19 (1.44x), Washington-Arlington-Alexandria $44.20 (1.32x), Seattle-Tacoma-Bellevue $44.13 (1.32x), Boston-Cambridge-Newton $43.09 (1.28x), New York-Newark-Jersey City $41.50 (1.24x), Denver-Aurora-Centennial $39.28 (1.17x), Atlanta-Sandy Springs-Roswell $34.57 (1.03x), Chicago-Naperville-Elgin $34.42 (1.03x, May 2024), Dallas-Fort Worth-Arlington $33.96 (1.01x), Phoenix-Mesa-Chandler $33.48 (1.00x). We damp the top end of those raw ratios. All-occupation means exaggerate the gap for the careers on this site, because national pay bands, remote hiring and company-wide equity grids compress geographic spread for high-skill professional roles more than they do for service and hourly work. Levels.fyi shows the same damping: its Bay Area software engineer average total compensation of about $291k sits roughly 1.3x-1.4x the US median, not 1.7x. Non-US multipliers convert local market rates to USD and are directional, not survey-grade.
How to become a Machine Learning Engineer
Every route we could verify, with honest time, cost and difficulty.
There are 5 routes we could verify into Machine Learning Engineer work. The fastest is Data engineer to ML engineer at about 18 months; the cheapest is Data scientist to ML engineer at about $800. Almost every Machine Learning Engineer hire in the United States comes from someone already employed as a software engineer, data engineer or data scientist who moved across after two to four years, usually inside the same company, because the interview loop tests production software engineering at the same bar as a backend engineer and that skill is built by doing the job rather than by coursework.
Software engineer to ML engineer
The dominant path and the one to plan for if you do not have a CS degree. Get any backend or data engineering job, become genuinely good at systems and testing, then move onto a team that owns models - often internally, by taking the ML-adjacent work nobody wants. Interview loops weight coding and system design heavily, so the engineering half is the half you cannot fake.
Data scientist to ML engineer
Works if you deliberately close the engineering gap: testing, code review, containers, CI/CD, distributed training, serving and latency. Many data scientists stall here because notebook habits do not survive the coding round. Budget 12-18 months of focused engineering work on top of an existing DS job.
Data engineer to ML engineer
Underrated. Data engineers already own pipelines, orchestration and warehouse performance, which is more than half of production ML. Add modelling, evaluation and serving and the transition is often a lateral move inside the same company with a significant pay rise.
MS in computer science or machine learning
Still the standard credential for new entrants and the most reliable way in for someone whose bachelor's is unrelated. Georgia Tech OMSCS (about $7,000-$9,000 total, part-time over 2-3 years) and UT Austin MSCSO (about $10,000) are the high-value options; full-time campus programmes cost $50,000-$120,000 and buy you internship recruiting access, which is the actual product.
Direct self-taught entry
Possible but rare, and you should plan around it being rare. It requires a genuinely strong public artefact - a widely-used open-source contribution, a serious Kaggle result, or a deployed system with users - plus the ability to pass a standard software engineering coding loop. Most people who try this end up taking a software or data engineering job first, which is the sensible outcome anyway.
Machine Learning Engineer roadmap: 8 steps, 1,360 hours
Becoming a Machine Learning Engineer from zero takes about 1,360 study hours across 8 steps, roughly 36 to 60 months at 10-15 hours a week plus a job search. Step one is Become a competent software engineer first. The roadmap puts production software engineering before mathematics and before modelling because the coding round is the stage that eliminates most career changers, and because every later step - data pipelines, MLOps, serving - is written in that same production code. Specialisation and portfolio systems come last, once you have something to specialise with.
0 of 8 steps done · saved in this browser ·
-
1
Python beyond scripting: modules and packaging, type hints, pytest, logging, profiling, virtual environments. Git branching and code review. The command line, SSH, and enough Linux to debug a container. Then data structures and algorithms to interview standard - arrays, hash maps, trees, graphs, two pointers, dynamic programming.
Why now: This is the step career changers skip and the reason they fail ML engineer loops. Every major employer runs the same coding round as for a backend engineer - Meta's ML engineer loop, for example, opens with two medium data-structures-and-algorithms problems before any ML question. No amount of modelling knowledge compensates for failing it.
- practice Top Interview 150LeetCode · 120 h · Free; Premium about $35/month
- book Fluent Python, 2nd EditionO'Reilly (Luciano Ramalho) · 45 h · About $80 print, or an O'Reilly subscription
- course The Missing Semester of Your CS EducationMIT · 15 h · Free
-
2
Linear algebra (vectors, matrices, eigen-decomposition, what a gradient is geometrically), calculus for backpropagation, probability and statistics. Then the classical algorithms and, more importantly, evaluation: cross-validation, leakage, calibration, class imbalance, precision-recall tradeoffs, baselines.
Why now: You need enough mathematics to reason about why a model is not converging and enough evaluation discipline not to ship a leaked feature. Andrew Ng's specialization is the standard grounding because it builds intuition before libraries; skip it only if you already have a quantitative degree.
- course Machine Learning SpecializationCoursera (Stanford Online and DeepLearning.AI) · 95 h · $49/month or included in Coursera Plus
- course Mathematics for Machine Learning SpecializationCoursera (Imperial College London) · 70 h · $49/month or included in Coursera Plus
- book Hands-On Machine Learning with Scikit-Learn, Keras, and TensorFlow, 3rd EditionO'Reilly (Aurelien Geron) · 60 h · About $80 print, or an O'Reilly subscription
-
3
Neural network fundamentals, backpropagation, optimisers, regularisation, batch norm, CNNs, RNNs and transformers. Then PyTorch specifically: datasets and dataloaders, writing a training loop, mixed precision, checkpointing, and multi-GPU training with DDP. Fine-tune a pretrained model on your own data at least twice.
Why now: Transformers underpin almost everything in modern ML, and PyTorch is the industry default. fast.ai gets you building working models fastest; the DeepLearning.AI specialization gives you the theory that interviewers ask about. Doing both in that order works well.
- course Practical Deep Learning for Coders (Part 1)fast.ai · 60 h · Free (9 lessons, about 90 minutes each, plus free book)
- course Deep Learning SpecializationCoursera (DeepLearning.AI) · 129 h · Free to audit; $49/month or Coursera Plus for certificate
- course Hugging Face LLM CourseHugging Face · 35 h · Free
-
4
SQL to fluency, warehouse modelling, batch pipelines with Airflow or Dagster, Spark or Ray for data that does not fit in memory, streaming basics with Kafka, and the concept of a feature store and why training/serving skew happens.
Why now: Surveys of ML practitioners consistently find that most production failures are data failures, not model failures. In small companies you will do this work yourself; in large ones you will be judged on whether you understand it. It is also the cheapest way to make yourself hireable while you build the rest of the skill set.
- cert Databricks Certified Data Engineer AssociateDatabricks · 45 h · $200 exam fee
- book Designing Data-Intensive ApplicationsO'Reilly (Martin Kleppmann) · 50 h · About $65 print
- course Data Engineer in Python career trackDataCamp · 55 h · Included in DataCamp Premium, about $28/month billed annually
-
5
Experiment tracking (MLflow or Weights & Biases), a model registry, containerising training and inference, a CI/CD pipeline that retrains and redeploys, an inference service with a latency budget, and monitoring for data and concept drift with an alert and a rollback path. Do it on real cloud infrastructure and pay the $30 bill.
Why now: This is the single clearest differentiator in an ML engineer interview: the 2026 loops explicitly probe deployment, CI/CD for models, drift handling, orchestration and endpoints. A candidate who has personally operated a model - including watching it degrade - answers those questions differently from one who has read about it.
- book Designing Machine Learning SystemsO'Reilly (Chip Huyen) · 35 h · About $65 print, or an O'Reilly subscription
- course Made With ML - MLOps courseGoku Mohandas · 50 h · Free
- cert AWS Certified Machine Learning Engineer - AssociateAmazon Web Services · 80 h · $150 ($75 during the MLA-C02 beta)
-
6
Pick one and go deep enough to hold an hour-long conversation: recommendation and ranking systems, computer vision, NLP and LLM systems, fraud and risk, forecasting and pricing, or ML platform/infrastructure. Learn the standard architectures, the standard metrics, and the standard failure modes of that domain.
Why now: ML engineer job descriptions are written by teams with a specific problem, and generalist applications lose to specialists. Ranking and recommendations has the most open roles and the most transferable interview content; ML platform is the least crowded and pays comparably.
- book Machine Learning System Design InterviewAli Aminian and Alex Xu · 30 h · About $40
- practice Papers with Code - track current baselines in your chosen domainPapers with Code · 40 h · Free
-
7
Not notebooks. System one: an end-to-end pipeline that ingests data on a schedule, retrains, evaluates against a baseline with a gate, deploys behind an API and reports drift on a dashboard. System two: something in your chosen domain with a real serving constraint - a ranking service that responds in under 100ms, or a fine-tuned model served with quantisation. Write a design document for each.
Why now: The design document is the artefact that makes a self-taught candidate credible, because it shows the tradeoff reasoning the ML design interview tests. Open-source contributions to PyTorch, Ray, vLLM, scikit-learn or Hugging Face libraries serve the same purpose and carry more weight with hiring managers.
- practice Kaggle competitions for a verifiable modelling signalKaggle · 60 h · Free
- practice Contribute to an open-source ML library (PyTorch, Hugging Face, Ray, vLLM)GitHub · 60 h · Free
-
8
150-250 LeetCode problems with an emphasis on mediums. Fifteen to twenty ML system design questions practised out loud on a whiteboard: design a news feed ranker, a fraud detector, a search relevance system, an ad CTR model. Deep-dive preparation on your own projects. Increasingly, an AI-assisted coding round where you are judged on how you direct the tool.
Why now: The standard 2026 loop is a recruiter screen, a technical screen of two medium algorithm problems, then four onsite rounds - coding, AI-assisted coding, ML design and behavioural. ML design is where most candidates underperform: they describe a model instead of a system with data collection, features, training cadence, serving, evaluation and monitoring.
- book Machine Learning System Design InterviewAli Aminian and Alex Xu · 35 h · About $40
- practice Top Interview 150 and the Grind 75 listLeetCode · 120 h · Free; Premium about $35/month
- book Designing Machine Learning Systems (chapters on evaluation, deployment and monitoring)O'Reilly (Chip Huyen) · 20 h · About $65 print
Best certifications for a Machine Learning Engineer
Which ones matter, what they cost, and how often people pass.
No certification is required for a Machine Learning Engineer role and no employer screens on one. The credential with real signal is the Certified Kubernetes Administrator at $445, because it is purely performance-based on live clusters, cannot be passed from question dumps, and Kubernetes is where serious training and inference workloads run - note that it expires after two years and renewal means sitting the whole exam again at full price. The AWS Certified Machine Learning Engineer - Associate at $150 and the Coursera specializations from Andrew Ng and DeepLearning.AI are curricula worth the study time, not credentials anyone hires on.
AWS Certified Machine Learning Engineer - Associate (MLA-C01, succeeded by MLA-C02)
Amazon Web Services
- Cost
- $150 USD for MLA-C01; $75 USD during the MLA-C02 beta
- Study
- 60-100 hours of prep with about a year of hands-on SageMaker and Bedrock experience
- Pass rate
- Not published. MLA-C01 is 65 questions in 130 minutes; MLA-C02 is 85 questions in 170 minutes. The older AWS Certified Machine Learning - Specialty (MLS-C01) was retired on 31 March 2026; MLA-C01's last English testing day is 28 September 2026.
Databricks Certified Machine Learning Associate
Databricks
- Cost
- $200 registration fee plus any local tax; all seven Databricks exams are $200 as of 2026
- Study
- 40-60 hours with prior Spark and Python experience
- Pass rate
- Not published
Deep Learning Specialization
DeepLearning.AI (via Coursera)
- Cost
- Free to enrol and audit; paid subscription ($49/month, or Coursera Plus at $59/month or $399/year) for graded assignments and the certificate
- Study
- 129 across 5 courses, about 3 months at 10 hrs/week
- Pass rate
- No exam; graded programming assignments. 4.8/5 from 147,248 reviews, over 1,000,000 enrolled. ACE recommendation of 10 college credits.
Machine Learning Specialization (Andrew Ng)
Stanford Online and DeepLearning.AI (via Coursera)
- Cost
- $49/month, about 2 months at 10 hrs/week; or included in Coursera Plus; financial aid available
- Study
- 95 across 3 courses (33 + 34 + 28)
- Pass rate
- No exam; 4.9/5 from 39,344 reviews
Databricks Certified Data Engineer Associate
Databricks
- Cost
- $200 registration fee plus local tax
- Study
- 40-50 hours
- Pass rate
- Not published
Skills employers screen for
Soft skills that decide offers: Scoping an ML problem, including deciding not to use ML, Writing design documents that survive review by senior engineers, Debugging systems where the failure is statistical rather than a stack trace, Working with product on what the metric should actually be, Communicating uncertainty and expected impact before a project is funded.
Best courses for a Machine Learning Engineer
Checked on the provider's page on 16 September 2026. Some links are affiliate links.
We list 6 courses for Machine Learning Engineer work, checked on the provider's page. Take one structured course for the theory - the Machine Learning Specialization at 95 hours or the Deep Learning Specialization at 129 hours, both $49 per month on Coursera - and then spend at least twice as many hours building and operating something real. The courses give you the vocabulary interviewers use; only the two production-grade portfolio systems in step 7 give you answers for the machine learning system design round.
Browse more Machine Learning Engineer courses on Coursera Coursera Plus covers most of the courses above for one monthly fee.
Machine Learning Engineer interview questions and format
A Machine Learning Engineer loop runs five to six stages over four to eight weeks: a recruiter screen, a technical screen of typically two medium data-structures-and-algorithms problems, then an onsite of four to five rounds covering coding, increasingly AI-assisted coding, machine learning system design, machine learning depth and behavioural. The machine learning system design round filters the most candidates, because people describe a model instead of a system with data collection, features, training cadence, serving, evaluation and monitoring.
Five to six stages over 4-8 weeks. (1) Recruiter screen. (2) Technical screen: typically two medium data-structures-and-algorithms problems in 45-60 minutes, sometimes preceded by an online assessment. (3-6) Onsite loop of four to five rounds: a coding round at the same bar as a backend engineer; increasingly an AI-assisted coding round where you use a code assistant and are judged on how you direct and verify it; an ML system design round (45-60 minutes, open-ended, end-to-end); an ML depth or breadth round on modelling, evaluation, and modern topics such as transformers, RAG and fine-tuning; and a behavioural round. MLOps is now explicitly tested - deployment, CI/CD for models, data quality and drift, pipelines and orchestration, and serving endpoints. Startups more often substitute a take-home of 4-8 hours plus a review session.
Questions that come up
- Design an end-to-end system to recommend items on a home feed for 100 million daily users. Cover data, features, training cadence, serving, evaluation and monitoring.
- How would you detect that a deployed model has degraded, and what would you automate?
- Explain training/serving skew, give a concrete example, and say how you would prevent it.
- Your offline AUC improved but the online A/B test is flat. What are the likely causes?
- Walk through backpropagation for a two-layer network, then explain what vanishing gradients are and how modern architectures avoid them.
- You have a 1-billion-row training set that does not fit in memory. How do you train?
- Explain attention and why transformers replaced RNNs for sequence modelling.
- How do you choose between a gradient-boosted tree model and a neural network for tabular data?
- Your inference p99 latency is 300ms and the budget is 100ms. What do you try, in order?
- Coding: given a list of intervals, merge the overlapping ones. (Standard medium - expect two of these.)
Prep
Machine Learning Engineer FAQ
Can I become an ML engineer without a CS degree?
Yes, but almost never as your first job. The realistic sequence is: get a software engineering, data engineering or data science role on whatever credential you can, spend two to four years becoming genuinely strong at production engineering, then move into ML - often internally. The reason is structural: the interview loop tests software engineering at full backend-engineer standard, and that skill is built by doing the job, not by coursework. An online MS such as Georgia Tech OMSCS (about $7,000-$9,000) is a cost-effective way to add the credential while you work.
How different is this from being a data scientist?
A data scientist is measured on whether the analysis or model produced a correct, useful answer. An ML engineer is measured on whether a system runs reliably, cheaply and fast, and keeps working when the data changes. In practice: data scientists write more SQL and statistics, ML engineers write more production code, infrastructure and tests. The pay gap is real - roughly $190,000 versus $280,000 at the Levels.fyi US median - and it is paid for the engineering, not the modelling.
Which machine learning certification is actually worth the money in 2026?
None of them get you hired on their own; senior engineers do not weight certificates highly. The two with practical value are the AWS Certified Machine Learning Engineer - Associate ($150, or $75 during the MLA-C02 beta) and the Databricks ML Associate ($200), because the exam content forces you to learn a real managed stack end to end and many enterprises use one of them. Note that the older AWS Machine Learning - Specialty (MLS-C01) was retired on 31 March 2026, so do not buy prep material for it. The Coursera specializations from Andrew Ng and DeepLearning.AI are worth the time for the knowledge, not the certificate.
Is the $280,000 median total compensation for machine learning engineers realistic?
It is the real median of Levels.fyi's US machine learning engineer submissions, but Levels.fyi is heavily weighted toward large tech employers and toward people who chose to report a good package. Glassdoor's broader sample for the same title averages $164,714 with a 90th percentile of $254,003. The honest reading: at a public tech company or a funded AI startup, $250,000-$400,000 at mid to senior level is normal; at a bank, insurer, retailer or mid-size SaaS company, $150,000-$210,000 is normal for the same title and often the same work.
Do I need to know deep learning, or is XGBoost enough?
For a great many production problems - fraud, churn, credit risk, pricing, demand forecasting - gradient-boosted trees are still the right answer and knowing why is a sign of maturity. But you cannot pass an interview in 2026 without being able to explain transformers, attention, fine-tuning and the basics of serving large models, because those questions are asked regardless of what the team actually runs. Learn both; default to the simple model in practice.
Machine learning engineer or AI engineer - which should I aim at in 2026?
ML engineer if you enjoy systems, infrastructure and training models, and you are willing to spend years building a software engineering base; it pays more at the top and the skills are more durable. AI engineer if you want to build products on top of existing foundation models and want a faster route in - the bar is closer to a strong application developer's, and the entry-level market is more open. Many people do AI engineering first and move into ML engineering later.
How many LeetCode problems do I actually need for a machine learning engineer loop?
Plan on 150-250 with the emphasis on mediums, spread over three to four months rather than crammed. Pattern recognition matters more than volume: arrays and hashing, two pointers, sliding window, binary search, trees and graph traversal, heaps, and basic dynamic programming cover most of what appears. Expect two medium problems in a 45-minute screen, and expect to be asked to explain your complexity.
What does a Machine Learning Engineer do all day?
Mostly engineering, not modelling. A typical day opens by checking an overnight training run that failed, often because a schema change upstream broke the feature pipeline, then fixing it, adding an assertion so it fails loudly next time, and opening the pull request. The middle of the day is design review, code review and profiling - finding that tokenisation is 40 percent of p99 latency and batching it. The end of the day is launching a shadow deployment and writing the weekly update. On-call means a model returning garbage at 2am because an upstream feed changed format: roll back first, diagnose after.
Is machine learning engineering a good career for someone switching at 40?
Yes as a second move, no as a first one. Nobody is hired into a Machine Learning Engineer seat directly from an unrelated career, because the loop tests production software engineering at the same bar as a backend engineer. The workable plan at 40 is the same as at 25: take a software, data engineering or data science job first, then move across in two to four years. Age matters less than the sequence does, and domain knowledge from a prior career is genuinely valuable in fraud, healthcare or industrial machine learning teams. Georgia Tech OMSCS costs about $7,000 to $9,000 and can be done while working.
How much does a machine learning engineer make in the San Francisco Bay Area versus the national median?
About 35 percent more on the offer letter. Levels.fyi puts the United States machine learning engineer median total compensation at $280,000; applying the Salary Roadmap San Francisco Bay Area multiplier of 1.35 gives roughly $378,000, and San Jose at 1.42 gives about $398,000. That spread is narrower than raw cost-of-living tables imply, because company pay grids compress geography for high-skill roles. It is also spent locally: California income tax and Bay Area rent absorb most of the premium below the senior rung, so compare after-tax income minus housing before you move.
How long does it take to become a machine learning engineer while working full time?
Three to five years, not months. The roadmap on this page totals about 1,360 study hours, which is roughly two years at 10 to 15 hours a week, but study is not the binding constraint - the two to four years of paid production engineering that make you credible in the loop are. The fastest honest sequence is to reach a software or data engineering job within 12 to 18 months, then close the modelling and MLOps gap on the job and move internally. People who try to go straight from a course to the title almost always end up taking the engineering job anyway.
Sources
Every number on this page traces to one of these. Page checked 16 September 2026.
- levels.fyi/t/software-engineer/title/machine-learning-engineer/locations/united-states
- levels.fyi/companies/meta/salaries/software-engineer/title/machine-learning-engineer
- levels.fyi/companies/google/salaries/software-engineer/title/machine-learning-engineer
- levels.fyi/companies/amazon/salaries/software-engineer/title/machine-learning-engineer
- levels.fyi/companies/apple/salaries/software-engineer/title/machine-learning-engineer
- levels.fyi/companies/nvidia/salaries/software-engineer/title/machine-learning-engineer
- glassdoor.com/Salaries/machine-learning-engineer-salary-SRCH_KO0,25.htm
- bls.gov/ooh/computer-and-information-technology/software-developers.htm
- bls.gov/ooh/computer-and-information-technology/computer-and-information-research-scientists.htm
- bls.gov/ooh/math/data-scientists.htm
- aws.amazon.com/certification/certified-machine-learning-engineer-associate/
- databricks.com/learn/certification
- coursera.org/specializations/deep-learning
- coursera.org/specializations/machine-learning-introduction
- course.fast.ai/
- huggingface.co/learn
- coursera.org/courseraplus