Every Tuesday I get the same email from somebody halfway through a career change. It reads something like, “I want to break into machine learning. Do I need a PhD? Should I do a bootcamp? Am I too late?” After fifteen years of recruiting machine learning talent, I can tell you the answers are no, probably not, and absolutely not.
This is the guide I send to those emails. It covers the four real paths into ML work, the skills that show up in every hiring conversation I have, and what the 2026 job market actually pays for each experience level.
The four paths into ML (and which one hiring managers prefer)
There are four legitimate routes into ML engineering, and all four are represented on every ML team I place into. In rough order of frequency I see them on hired resumes:
- Computer science or statistics degree with ML coursework, followed by internships and early-career roles
- Adjacent engineering background (software, data, backend) with self-directed learning and a portfolio
- Graduate degree (master’s or PhD) in ML, statistics, physics, or a quantitative field
- Bootcamp or structured online specialization, paired with significant self-driven project work
No hiring manager I work with will disqualify a candidate for the path they took. Every hiring manager I work with will disqualify a candidate who cannot talk through a real project they shipped.
The core technical skills that show up in every job description
Ignore the job description arms race where every posting lists twenty frameworks. The core stack that actually shows up in interviews is narrow:
- Python: fluent, not conversational. You should be able to build a small service, write tests, and debug someone else’s code
- PyTorch and the PyTorch ecosystem: TensorFlow is still around, but PyTorch dominates new work
- SQL and data manipulation: dataframes, joins, window functions. This is where a surprising number of candidates fall apart
- Classical ML fundamentals: linear and logistic regression, trees, gradient boosting, cross-validation, bias-variance
- Deep learning fundamentals: backpropagation, attention, transformers, embedding spaces
- Cloud basics: at least one of AWS, GCP, or Azure, with working knowledge of object storage, compute, and a managed ML service
- Git, testing, and basic software engineering hygiene: the thing that separates researchers from engineers
Do those well and you will beat most of the field. Try to do twenty of them poorly and you will lose to someone who did seven of them cleanly.
The soft skills that separate promotable engineers from plateaued ones
The engineers I see promoted inside eighteen months of joining a team share three traits. They can explain a model tradeoff to a product manager without condescending. They write documentation that their teammates actually read. They ask clarifying questions before they write a single line of code. None of those skills are on a typical ML curriculum, and all three are the reason one candidate gets the offer over an equally technical peer.
Your portfolio matters more than your resume
A resume tells a hiring manager where you have been. A portfolio tells them what you can actually do. The portfolios that land interviews for career switchers share a pattern: two or three substantial projects, each one with a real dataset, a shipped artifact (a working model, a deployed app, a public notebook), and a written explanation of what went wrong and how it was fixed. Kaggle competition entries count, but only if they include reflection, not just leaderboard screenshots.
I once placed an operations research analyst into a senior ML role based almost entirely on a supply-chain forecasting portfolio she had built on her own time. She had zero ML on her resume, and four projects on her GitHub. The hiring manager said, “I can see how she thinks.”
The 2026 job market: where the jobs are, where they are not
The shape of the market has shifted. Three categories of role are expanding the fastest in 2026:
- Applied ML engineers inside product organizations, ranking, recommendation, search, personalization, fraud
- ML platform and MLOps engineers, building the infrastructure the model teams ride on
- LLM and generative AI engineers, both research-adjacent and applied, and sitting in every company that was not an AI company two years ago
The categories that are growing more slowly are pure research roles outside of the frontier labs, and generalist data scientists whose work overlaps heavily with modern analytics tooling.
Salary expectations by region and experience level
Here is the honest salary picture for first-role ML candidates in 2026, base salary only, cross-checked against BLS, Levels.fyi, and placements I have closed in the last six months:
- Entry-level in a top metro (SF, NYC, Seattle, Boston): $135,000 to $175,000
- Entry-level in a secondary metro (Austin, Denver, Atlanta, Chicago): $115,000 to $150,000
- Remote-first companies paying metro rates: add roughly ten to fifteen percent to the secondary-metro band
- Entry-level with a PhD or top-program master’s: $165,000 to $225,000 base
Equity at private companies varies so widely that I do not include it in a table. Ask for specifics: strike price, vesting, and the last 409A valuation.
How to land your first role (without three years of experience)
The chicken-and-egg problem is real. Every job wants experience you cannot get without a job. The people I see break through do some combination of the following: build visibly, contribute to open-source projects that matter, take a lower-paying adjacent role that puts them close to the ML team, and network relentlessly with engineers who can recommend them in. Referrals are not nepotism in this market; they are the only efficient signal a hiring manager has.
Your first ML job is closer than you think
If you are reading this and wondering whether the door is closed, it is not. It has narrowed, and the criteria are sharper than they were three years ago, but the demand for clear thinkers who can ship models has never been higher. Pick a path, commit to the core stack, ship something you can show, and get yourself in front of the people doing the Machine Learning work you want to do. I have seen people go from zero to first offer in eighteen months. It is not magic; it is a plan.
