First, pick the right target
Data scientists explore data and prototype models to answer business questions. ML engineers take models to production: pipelines, feature stores, serving, monitoring, retraining. AI engineers build systems on top of foundation models: retrieval, agents, evaluation. The demand curve has bent toward the two engineering roles, because every company that prototyped models in the 2020s now needs someone to run them reliably.
This post is the ML engineering path. If LLM product systems excite you more, read the AI engineer version; the foundations overlap enough that you can switch.
The skill order that works
- Python, NumPy, pandas, statistics. The numerical toolkit, from basic arithmetic if needed.
- Exploration and feature engineering. Where most real model quality comes from.
- Supervised learning with honest baselines. Beat naive first; celebrate later.
- Evaluation discipline. Splits, leakage, and metrics matched to decisions: the professional differentiator.
- Ensembles and optimization. Gradient boosting still wins most tabular problems.
- Neural networks and PyTorch. Training loops as a debuggable machine.
- Applied tracks. Vision, NLP, forecasting, recommendations, mostly via adaptation.
- Pipelines and reproducibility. Rerun last month's experiment; get the same number.
- Serving, MLOps, monitoring, drift. Models as living systems.
- Scale and platforms. Distributed training, accelerators, responsible ML, LLM systems.
The stage-by-stage version with practice targets is in the ML engineer roadmap guide; the ML Engineering app teaches this exact sequence as 40 courses and 400 interactive lessons ending in a principal ML engineering capstone.
The mistake that costs the most time
Skipping classic ML to reach deep learning faster. Classic ML is where you learn the failure modes (leakage, overfitting, weak baselines, misleading metrics) at a scale where they are visible and cheap. Engineers who skip it ship neural networks that lose to a well-tuned boosted baseline, and cannot tell, because they never learned to build honest baselines. Every hiring manager in the field has seen this candidate; do not be them.
Practice like the job tests you
ML fluency decays into vocabulary without validated practice. The app's 400 lessons each end in a checked challenge: 140 editable code labs across NumPy, pandas, and PyTorch, plus 80 architecture decisions that rehearse the production judgment interviews probe: batch versus realtime serving, retrain versus monitor, bigger model versus better features. One honest hour daily, typed, beats a weekend of videos.
The portfolio that proves engineering
- A tabular project with an honest baseline. Boosted trees versus alternatives, leakage checks documented.
- An adaptation project. A pretrained vision or NLP model adapted to a new task with defensible evaluation.
- A reproducible pipeline. Data to model to metrics, rerunnable end to end from a fresh clone.
- A serving and monitoring design. Deployment pattern, quality monitoring, drift detection, retraining plan.
The app's Project Lab covers this gradient across 40 briefs with editable workspaces and milestones. Rebuild your best ones publicly with README trade-off discussions; that writing is what separates engineers from notebook users in a reviewer's first minute.
A realistic timeline
From zero, eight to fourteen months of daily practice to a credible junior ML engineering portfolio; four to eight from a software base. Data analysts land in between, converting statistics knowledge into engineering habits.
Where to start today
If you want the whole path in one place, ML Engineering is a native Mac app with the full sequence, offline validated practice, a private curriculum-grounded tutor, and progress that stays on your machine, available as a donation download on this site.
FAQ
ML engineer, data scientist, or AI engineer?
Pick by layer: exploration, productionizing models, or building on foundation models. Demand favors the engineering roles.
Do I need a PhD?
Only for research. Production ML engineering hires on demonstrable engineering skill.
What goes in the portfolio?
An honest tabular baseline, an adaptation project, a reproducible pipeline, and a serving-plus-monitoring design, all documented.
How long?
Eight to fourteen months from zero with daily practice; less from adjacent roles. Classic ML first; it is the fast path, not the slow one.