AI engineer roadmap: what to learn and in what order
An AI engineer roadmap across eight skill domains: what to learn first, what each role needs, and what you can skip for the job you want.
On this page
An AI engineer roadmap has one rule: learn in dependency order, and only to the depth your target role needs. Start with Python, Git and SQL, add data handling and the maths your role requires, then machine learning fundamentals and evaluation, then deep learning and embeddings, then LLM skills such as retrieval, agents and evaluation, and finish with deployment and system design. The order is fixed by prerequisites. The depth is decided by the role.
That second point is where most plans go wrong. People study everything to the same depth, run out of energy somewhere in calculus, and never reach the skills their target job description actually lists.
What does an AI engineer need to learn?
An AI engineer needs skills from eight domains, and no single role needs all eight in depth. The domains below are the ones this site's concept map uses to describe the seven AI engineering roles at GCCs.
| Domain | What it covers | Core for |
|---|---|---|
| Maths and statistics | Linear algebra, calculus, probability, inference, experiments, optimisation | ML engineer, GenAI engineer, applied scientist |
| Programming and computer science | Python, Git, SQL, data structures, testing, APIs, async, system design | Every role |
| Data | Modelling, pipelines, distributed processing, data quality, feature engineering, governance | AI data engineer, then ML and GenAI engineers |
| Classical machine learning | Fundamentals, supervised learning, model evaluation, common libraries | ML engineer, GenAI engineer, applied scientist |
| Deep learning | Neural networks, a training framework, embeddings, transformers | ML engineer, GenAI engineer, applied scientist |
| LLMs and generative AI | Prompting, LLM APIs, vector search, RAG, agents, evaluation, safety | GenAI engineer, ML engineer, AI backend engineer, AI architect |
| MLOps and infrastructure | Containers, cloud, CI/CD, deployment, pipelines, monitoring | AI platform and MLOps engineer, ML engineer |
| Systems and leadership | Communication, problem framing, ML and GenAI system design | AI architect, then senior levels of every role |
Three skills are core for all seven roles: Python, system design and cloud platforms. Git, testing and containers are core for six of the seven. If you are unsure where to begin, those are never wasted.
In what order should you learn AI engineering skills?
Learn AI engineering skills in the order their prerequisites dictate: foundations, then data and maths, then machine learning, then deep learning, then LLMs, with deployment skills running alongside from the middle onwards. Each stage below depends on the one before it.
Programming foundations
Python, Git and the command line, SQL. Then data structures, testing and APIs, which need Python first.
Data handling and maths
NumPy and Pandas, probability and statistics, linear algebra. Add calculus and optimisation if your role trains models.
Machine learning fundamentals
How models generalise, supervised learning, and model evaluation. Evaluation is the prerequisite for most later stages.
Deep learning and embeddings
Neural networks, a training framework, embeddings, then sequence models and transformers.
LLM applications
LLM fundamentals, prompting, LLM APIs, vector search, RAG, agents, then safety and LLM evaluation.
Production
Containers, cloud, CI/CD, experiment tracking, deployment and monitoring. Start containers early; the rest needs evaluation first.
System design
ML system design and GenAI architecture, which need deployment, evaluation, RAG and agents behind them.
A few chains inside that order are worth memorising, because skipping a link is the usual reason a topic feels impossible:
- The retrieval chain. Neural networks, then embeddings, then vector search, then RAG, then agents, then guardrails. You can call a RAG library without the first two links, but you cannot debug it.
- The evaluation chain. Statistics and linear algebra, then ML fundamentals, then model evaluation, then LLM evaluation and experiment tracking.
- The deployment chain. Git, then containers, then cloud. Testing plus containers, then CI/CD. APIs, containers and experiment tracking, then model deployment, then monitoring.
The individual stages each have their own guide: the Python an AI engineer actually uses, SQL for AI and ML roles, how much maths machine learning needs, machine learning fundamentals, how RAG works, AI agents and tool use, evaluating an LLM application and MLOps basics.
How does the roadmap change by role?
The roadmap changes by role in depth, not in order: every role walks the same sequence but stops early in some domains and goes deep in others. Pick the row below that matches your target and treat "skip for now" literally.
| Target role | Go deep on | Working knowledge | Skip for now |
|---|---|---|---|
| ML engineer | Maths, classical ML, deep learning, deployment, monitoring | SQL, data pipelines, statistics for experiments | Infrastructure as code |
| GenAI and LLM engineer | LLM stack, embeddings, transformers, evaluation, deployment | SQL, classical ML breadth | Infrastructure as code |
| AI platform and MLOps engineer | Containers, Kubernetes, CI/CD, serving, pipelines, monitoring | ML fundamentals, LLM APIs, vector search | Maths, most classical ML |
| AI data engineer | SQL, data modelling, pipelines, streaming, data quality, vector search | ML fundamentals, statistics, RAG | Calculus, deep learning theory |
| AI backend engineer | APIs, async, system design, LLM APIs, RAG, agents, safety | ML fundamentals, evaluation, embeddings | Maths, model training |
The two remaining roles sit at the ends. A data or applied scientist goes deepest on statistics, experiments and modelling and treats infrastructure as working knowledge. An AI architect or lead needs working knowledge of nearly everything and depth in system design, cost, governance and build-versus-buy decisions, which is why that role follows years in one of the others.
For the detail of each track, read about the GenAI and LLM engineer role, the MLOps and AI platform engineer role and the AI backend engineer role, or compare all seven in the role explorer.
Where should a working software engineer start?
A working software engineer should start at machine learning fundamentals and evaluation, not at Python or maths. If you already write tested services, use Git daily and can design an API, the first stage is done, and the fastest routes are the AI backend and AI platform roles, which do not treat maths as a requirement.
Student or fresher
- Recent maths, so the theory stages move quickly
- Spend the most time on Git, testing, APIs and SQL
- Deploy at least one project end to end
- Target roles that train and evaluate models
Service-company engineer
- Engineering habits already in place
- Start at ML fundamentals and evaluation
- Add the LLM stack: prompting, retrieval, agents
- Target AI backend, platform or data roles first
The software engineer to AI engineer transition plan covers that route in more detail.
What are the most common roadmap mistakes?
The most common roadmap mistake is studying without a target role, which makes every topic look equally important. Four others follow from it:
- Stopping at tutorials. A topic is learned when you have built something with it and measured the result. Each stage should end with a small project; see AI portfolio projects that get interviews.
- Skipping evaluation. Candidates who can build a model or a RAG pipeline but cannot say how they know it works are filtered out in machine learning interviews and GenAI interviews alike.
- Treating deployment as optional. Containers, cloud and monitoring are core for most roles. A model that only runs in a notebook is half the work.
- Leaving system design until the interview is booked. Design questions draw on everything earlier in the roadmap, so practise them as you go using a framework for ML system design interviews.
How do you know where you stand?
You know where you stand by checking yourself against the core skills for one role, concept by concept, and being strict about the difference between "I have read about it" and "I could explain it and build it". A gap list that short is far more useful than a general feeling of being behind.
You can do that by hand with the tables above. If you would rather have it done for you, the free readiness check asks which role you want, scores you against that role's concepts and turns the gaps into a personal roadmap in the right order.
Key takeaways
- Choose a target role first: the same eight domains apply to everyone, but the depth each role needs is very different.
- Python, system design and cloud platforms are core for all seven AI roles, so no one skips them.
- Follow prerequisites, not hype: embeddings come before vector search, vector search before RAG, and RAG before agents.
- Evaluation is the skill that separates a demo from a system, for classical models and LLM applications alike.
- Backend, platform and data engineers can reach AI roles without deep maths; ML engineers, GenAI engineers and applied scientists cannot.
Frequently asked questions
Can I become an AI engineer without a machine learning degree?
Yes. Interviews test what you can build and explain, not which degree you hold. The roles closest to software engineering, such as AI backend engineer and AI platform or MLOps engineer, lean on skills many working engineers already have. A degree helps most for applied scientist roles, where statistics and research depth are examined directly.
Should I learn classical machine learning before LLMs?
Learn the fundamentals first, even if your goal is LLM work. Train and test splits, overfitting, baselines and metrics apply directly to evaluating an LLM application. You do not need every classical algorithm, but a GenAI engineer who cannot explain how to measure quality will struggle in interviews and on the job.
Do I need to learn deep learning frameworks to build LLM applications?
Not to start. Applications built on hosted models need Python, APIs, prompting, retrieval and evaluation more than a training framework. A framework such as PyTorch becomes necessary when you fine-tune models, debug embeddings or apply for ML engineer and GenAI engineer roles, where training knowledge is treated as core.
How do I know which parts of the roadmap I can skip?
Compare your target role against each domain. A domain marked not needed for the role can wait; a domain marked good to know needs working familiarity, not depth. Then test yourself honestly on the core items, because the parts people skip are usually the ones they assume they already know, such as SQL and testing.
Is the roadmap different for freshers and experienced engineers?
The order is the same, but the starting point differs. A fresher usually has recent maths and little production experience, so deployment, testing and system design need the most work. An experienced service-company engineer usually has the engineering habits and needs machine learning fundamentals, evaluation and LLM skills.