Machine Learning Engineer

What this work pays

The federal wage survey has no code for this title, so these are computer and information research scientists: median $140,300, bottom tenth $82,200, top tenth $230,630. That code leans toward research roles with advanced degrees. A lot of working machine learning engineers are actually counted as software developers instead, where the median is lower. Take this as the top of the software pay band rather than a typical first salary.

First job out of school

Location changes this enormously. The San Francisco Bay Area, Seattle and New York pay a long way above the rest of the country, and they also cost a long way more to live in.

Stock, not just salary

Published salary figures usually miss this. At big technology firms a meaningful part of total pay is company shares that vest over several years, which is also what keeps people from leaving.

What it costs to get there

Public university, in-state, four years

Roughly $11,000 to $15,000 a year in tuition and fees

Community college for the first two years

Commonly $4,000 to $9,000 in total, then transfer

Master's degree, if you want one

$20,000 to $80,000, though online options exist for far less

A laptop and cloud computing time

$1,500 to $3,000, and most learning platforms have free tiers

Roughly, all in

$1,500 to $90,000

Tuition, fees, exams and kit. Not rent, food or travel

The low end is the self taught route, free course material, open datasets and a few hundred dollars of cloud computing time, and people really do get hired that way. Be honest about it though, because that door almost always opens for someone already working as a software developer who is moving across, not for someone starting from nothing. The high end is a four year public university degree at in state rates plus a master's, which is the common path. Two years at a community college then transferring cuts the degree cost sharply, and internships in this field pay well enough to cover a slice of a year's tuition.

File the FAFSA (Free Application for Federal Student Aid). Pell Grants are not repaid, and federal loans for a dependent undergraduate cap at $5,500 the first year and $31,000 across the degree. Paid internships in this field are unusually well paid and can cover a large slice of a year's tuition on their own.

What comes with the job

• Remote and hybrid work

  Still more common here than almost any other well-paid field.

• Conference and training budgets

  Employers pay for courses and conferences because the field changes fast.

• Fast pay growth early

  The jump from the first job to the third is steeper here than in most careers.

• Skills that transfer across industries

  Hospitals, banks, farms, insurers and governments all hire for this now, not just technology companies.

The path

A degree plus proof you can actually build things. Both halves are required.

1. 1

   High school

   Grades 9 to 12

   Take the hardest math available and start writing code, ideally Python. Everything after this assumes you already enjoy both.

2. 2

   Bachelor's degree

   4 years

   Computer science, mathematics, statistics or a related engineering field. Community college for two years then transferring to a public university is the cheapest version of the same degree.

3. 3

   Build and ship things

   Throughout

   Personal projects, open source contributions, competition entries. A working project someone else uses beats a course certificate every time.

4. 4

   Internships

   Summers

   Apply from your second year. This is how most people get their first full-time offer, and many companies hire almost exclusively from their own interns.

5. 5

   First job, then decide on a master's

   0 to 2 more years

   Engineering roles rarely need one. Research roles usually want a master's or a doctorate. Working first tells you which you actually want.

High school courses that help

• Calculus

  Model training is built on derivatives. You will meet them again on day one.

• Statistics and probability

  Arguably more important than calculus, and far more often skipped.

• Computer science

  Any programming at all. Python is the language this field runs on.

• Physics

  Good practice at turning a messy real situation into a model you can actually compute with.

Time and money, at a glance

Years after high school

4 to 6

Admission

Competitive at strong computer science programs, open at many public universities

Total tuition

$20,000 to $60,000 in-state at a public university

Paid while training?

Internships pay very well, often $7,000 to $12,000 a month

Nobody licenses this, and that matters

There is no board, no exam and no legal title. Anyone can call themselves a machine learning engineer, and hiring is done almost entirely by technical interview and portfolio. That cuts both ways: it means there is no gatekeeper keeping you out, and it means the interview is genuinely hard. Rules are starting to arrive on the use side rather than the practitioner side. The National Institute of Standards and Technology publishes a voluntary framework for managing risk in artificial intelligence systems, and some states and sectors regulate specific uses like hiring tools or medical devices. None of it licenses you personally.

Where people study

• Large public universities

  In-state tuition at a state flagship is the best value route into this field and employers hire heavily from them.

• Specialist departments

  A handful of universities have dedicated machine learning departments, for example Carnegie Mellon University.

• Online master's programs

  Georgia Tech's online computer science master's is the best known and costs a fraction of an on-campus degree.

• Self-taught plus a portfolio

  It happens, mostly for people already working as software engineers. Breaking in this way with no degree and no job history is rare.

Specialties that pay

• Large language models

  Fine-tuning, evaluation and building products on top of them. The busiest corner of the field right now.

• Machine learning infrastructure

  Serving models reliably at scale. Less glamorous, extremely well paid, and always short of people.

• Computer vision

  Medical imaging, manufacturing inspection, self-driving. Deep and durable.

• Applied research

  The route that usually wants a doctorate, and where the top of the pay range sits.

Nice-to-haves

• Solid software engineering

  Testing, version control, code review. Models that cannot be deployed are worth nothing, and this is the most common gap in new graduates.

• Cloud experience

  The major cloud platforms all have free tiers and cheap certifications. Employers do read them.

• Writing clearly

  You spend a lot of time explaining to non-technical people why the model did something strange.

• A public portfolio

  Code someone can read and a project someone can try, on a public repository.

The title means different jobs at different companies

At most companies this role is software engineering with a modelling component: you take a model someone else trained, adapt it, wire it into a product and keep it working. At a small number of research labs it means inventing new methods and publishing papers, and those jobs want a doctorate. Both are called machine learning engineer. Read the actual job description rather than the title, because the day-to-day work and the qualifications expected are very different.

A typical day

Start with whether last night's training run finished and whether the numbers moved. Most of the morning is data: finding it, cleaning it, working out why a chunk of it is wrong. Then a code review, a meeting with the product team about what the model is supposed to do, and an argument about how to measure that. Afternoon might be writing an evaluation, or chasing why the deployed model got slower this week. Actual model architecture work is a small slice of the whole thing.

The honest trade-offs

Work life

Typical hours

40 to 50 a week, with crunch before launches

Remote work

Common

Physical demand

Low

Unionized

No

Similar paths

• Data ScientistTurn messy company data into numbers a boss can act on, using code and statistics.
• Software DeveloperYou write the thing, then spend most of your time working out why it does not do what you meant.
• Data AnalystTurn a company's messy data into answers people can act on. The most reachable job in the data field.

The idea is older than the computers

The first artificial neuron model was published in 1943, and Frank Rosenblatt built the perceptron, a machine that learned from examples, in the late 1950s. The theory sat mostly unused for decades because the hardware and the data to make it work did not exist yet.

Most of the job is plumbing

Practitioners consistently report that the majority of their time goes on collecting, cleaning and labelling data rather than designing models. If you dislike messy spreadsheets and missing values, this field will disappoint you quickly.

Almost nobody trains a big model from scratch

Training a large model costs millions of dollars in computing time, so the overwhelming majority of engineers build on top of models trained by someone else. The valuable skill is not building one from nothing, it is knowing how to adapt and evaluate one.

There is no license, only guidance

Anyone may use this title. The National Institute of Standards and Technology publishes a voluntary framework for managing risk in artificial intelligence systems, and specific uses such as medical devices or hiring tools carry their own rules, but no state licenses the person doing the work.