A complete, honest roadmap for the skills AI engineers actually use, from programming and math foundations through machine learning, deep learning, LLMs, building AI applications, data engineering, MLOps, evaluation, responsible AI, and scaling. It runs top to bottom, foundational to advanced, so you always know what comes next. Free to read, no signup required.
How to use this: each step below is collapsed. Tap one to expand its details, skill pills, and guidance (only one opens at a time). Work down the spine in order; each stage assumes the ones above it. The field moves fast, so anchor on the durable concepts and build real projects, because that’s what turns AI tinkering into engineering.
01
AI engineering is software engineering with AI at its core. Before models, get genuinely good at programming, because most of the job is still building reliable software.
Python, deeply: Python is the lingua franca of AI. Learn it properly: data structures, control flow, error handling, and idiomatic style.
Software engineering: Version control, testing, clean code, and working in a codebase. AI systems are software, and they rot without these.
APIs & the web: How services talk over HTTP, since most AI features are delivered as (and consume) APIs.
Data handling: Reading, transforming, and validating data in code: the everyday substrate of all AI work.
“AI engineer” spans a spectrum, from applying existing models in products to training them from scratch. Both ends rest on solid software engineering, which is why it comes first.
PythonGit & testingClean codeAPIsData handling
02
You don’t need a maths PhD, but you do need enough intuition to understand why models behave as they do. Aim for working understanding, not proofs.
Linear algebra: Vectors, matrices, and operations: the language models are built and computed in (embeddings are vectors).
Probability & statistics: Distributions, expectation, and uncertainty: the foundation of how models learn and how you evaluate them.
Calculus intuition: Gradients and optimisation at a conceptual level, enough to understand how models are trained.
Why it matters: The goal is intuition that helps you debug, evaluate, and reason about models, not memorising formulas.
Linear algebraProbabilityStatisticsCalculus intuitionOptimisation
03
Classic ML is the conceptual bedrock beneath modern AI. These ideas (training, generalisation, evaluation) apply to every model you’ll ever work with.
Supervised & unsupervised learning: Learning from labelled data vs finding structure in unlabelled data: the two core paradigms.
Training & generalisation: Train/validation/test splits, overfitting vs underfitting, and the bias-variance trade-off.
Evaluation metrics: Accuracy, precision/recall, and why the right metric depends entirely on the problem. A wrong metric hides a broken model.
Core algorithms: A working knowledge of workhorse models (regression, trees, clustering) that still solve many real problems better than deep learning.
Don’t skip classic ML to jump straight to LLMs. Overfitting, evaluation, and generalisation are exactly the concepts that separate people who ship reliable AI from people who ship demos.
Large language models are the centre of gravity in applied AI today. Learn to use them effectively, because this is where most AI engineering work now happens.
How LLMs work: Tokenisation, context windows, and the idea of predicting the next token, enough to reason about their strengths and failure modes.
Prompting: Designing prompts, giving examples, and structuring instructions to get reliable output. A real, learnable skill.
Embeddings: Turning text into vectors to measure similarity and power search and retrieval.
Adapt vs prompt: When prompting is enough versus when fine-tuning or other adaptation is worth the cost. A key judgement call.
TokenisationContext windowsPromptingEmbeddingsFine-tuning vs prompting
06
Turning a model into a product is its own discipline. This is the day-to-day of most applied AI engineering roles.
Retrieval-augmented generation (RAG): Answering from your own data by retrieving relevant context at query time. It’s the dominant pattern for factual, current AI features.
Vector databases: Storing and searching embeddings efficiently to power retrieval at scale.
Tool use & agents: Letting models call functions and tools to act, and orchestrating multi-step workflows with appropriate guardrails.
Structured output & reliability: Getting predictable, parseable output, handling failures, and designing around non-determinism.
Most production AI value today comes from applying models well (retrieval, orchestration, evaluation, and reliability) rather than training new ones. This step is where careers are being built right now.
RAGVector databasesTool use & agentsOrchestrationStructured output
07
Models are only as good as their data. A large share of real AI work is getting the right data, clean and in the right shape.
Data pipelines: Collecting, cleaning, and transforming data reliably and repeatably at scale.
Data quality & labelling: Ensuring data is accurate and representative, and how labelled datasets are created and maintained.
Preparing data for models: Chunking, embedding, and structuring data for training and retrieval.
Bias in data: Recognising that biased or unrepresentative data produces biased models. It’s a data problem before it’s a model problem.
Data pipelinesData qualityLabellingChunking & embeddingData bias
08
A model in a notebook helps no one. Getting models into production and keeping them healthy is what makes you an engineer, not merely an experimenter.
Serving models: Exposing models behind APIs and running inference reliably in production.
Versioning & reproducibility: Tracking models, data, and experiments so results can be reproduced and rolled back.
Monitoring & drift: Watching quality, latency, and cost in production, and detecting when a model degrades as the world changes.
CI/CD & containers: Automated pipelines and containerisation to deploy AI systems consistently.
Model servingVersioningMonitoringModel driftCI/CD & containers
09
Evaluating AI is harder than testing normal software because output isn’t simply right or wrong. Rigorous evaluation is a defining skill of a serious AI engineer.
Offline & online evaluation: Measuring quality on held-out data before launch, and monitoring real-world performance after.
LLM-specific evaluation: Building eval sets, using rubrics and model-graded evaluation, and tracking quality as prompts and models change.
Guardrails: Detecting and constraining bad output: hallucinations, unsafe content, and off-task responses.
Measuring what matters: Tying evaluation to real user value, not merely benchmark scores that can mislead.
Without solid evaluation you’re flying blind, unable to tell whether a change helped or hurt. Investing in evals is what turns AI tinkering into engineering.
AI systems affect real people, so building them responsibly is part of the engineering, not an afterthought. This is increasingly a core professional expectation.
Bias & fairness: Recognising and mitigating unfair or harmful behaviour that data and models can encode.
Privacy & security: Protecting user data, and defending against prompt injection and misuse wherever user input reaches a model.
Transparency: Being honest about what a system can and can’t do, showing sources, and keeping humans in control.
Governance: Awareness of the emerging norms and regulations around deploying AI responsibly.
Running AI at scale is an engineering challenge of its own. Cost, latency, and hardware become central once real traffic arrives.
Cost & latency: Managing the real trade-offs between quality, speed, and cost that dominate production AI decisions.
Optimising inference: Caching, batching, and other techniques to serve models efficiently at scale.
Hardware awareness: How GPUs and accelerators affect what’s feasible, and the basics of distributed training and serving.
Architecture: Designing systems where models are one component among many, with the same reliability care as any backend.
Cost vs latencyCaching & batchingGPUsDistributed computeSystem design
12
AI moves fast, but the fundamentals move slowly. The durable skill is learning continuously while telling real progress from hype.
Learning continuously: Following credible research and practitioners, and trying new techniques on real problems rather than chasing every headline.
Fundamentals over hype: Anchoring new tools in the fundamentals above, so you can judge what’s genuinely useful versus marketing.
Evaluate before adopting: Testing whether a new model or technique actually helps your use case before rolling it out.
Build in public: Shipping projects and sharing what you learn. It’s the fastest way to grow and to be seen in a young field.
Tools and model names on this page will change; the concepts (evaluation, retrieval, reliability, responsible deployment) will not. Anchor on the fundamentals and the tools become easy to pick up.
Continuous learningFundamentals firstEvaluating new toolsBuilding in public
Build and evaluate real AI projects
AI engineering is learned by building, and just as importantly by evaluating what you build. A couple of real projects that you can measure and reason about prove far more than any certificate in a field this new.
A RAG app over your own documents, with a real evaluation set to measure answer quality.
A small agent that uses tools to complete a multi-step task, with guardrails and error handling.
Fine-tune or adapt a smaller model for a focused task, and measure whether it actually beats prompting.
Take one project to production: serve it behind an API, monitor cost and latency, and track quality over time.
Put it on GitHub with a README that reports how you evaluated it. The eval is what shows you’re an engineer, not merely a prompt tinkerer. Being able to say why your system works, and how well, is the whole game.
Frequently asked questions
What’s the difference between an AI engineer, an ML engineer, and a data scientist?
The lines are blurry and vary by company. Broadly: data scientists lean toward analysis and modelling to answer questions; ML engineers focus on training and productionising models; AI engineers increasingly focus on building applications on top of foundation models (LLMs): RAG, agents, evaluation, and reliable delivery. This roadmap covers the fundamentals under all three so you can move between them.
Do I need a PhD or a machine-learning degree?
No, especially for the fast-growing applied end of AI engineering, which is about building reliable software on top of existing models. A strong grasp of the fundamentals here, plus real projects, matters more than a specific degree. Research roles that train new models from scratch are where advanced degrees are more common.
How much math do I really need?
Enough for intuition, not proofs. Linear algebra (vectors and matrices), probability and statistics, and a conceptual grasp of gradients and optimisation are what help you evaluate and debug models. You can go a long way in applied AI engineering with working intuition rather than deep theoretical mathematics.
Should I start with LLMs or classic machine learning?
Build the foundation first. It’s tempting to jump straight to LLMs, but concepts like evaluation, overfitting, and generalisation come from classic ML and are exactly what separate reliable AI systems from impressive demos. Learn the fundamentals, then apply them to LLMs.
How long does it take to become an AI engineer?
It depends heavily on your starting point and how much time you invest, and far more on real projects than any fixed number. Someone with a strong software background can move into applied AI engineering faster than someone starting from zero programming. Building and evaluating real AI projects is what accelerates it.
Won’t the specific tools be outdated soon?
Some will. The field moves quickly, which is exactly why this roadmap emphasises durable concepts (evaluation, retrieval, reliability, responsible deployment) over any single tool or model. Anchor on the fundamentals and new tools become easy to learn as they arrive.
Early access
AI Engineer support is coming to Interview Ready
Our personalized Interview Ready plans don’t cover AI Engineer just yet, so unlike our other roles, this page doesn’t hand you off to a plan we can’t build for you honestly. This roadmap is free to use in the meantime. Want to be first in line when we add AI Engineer? Register your interest and we’ll let you know the moment it’s ready.
