AI Engineer roadmap
Build dependable products on top of language models: prompts, retrieval, tools, evaluation and safe operation.
At a glance
- Stages6
- Steps23
- Core steps15
- Optional or alternative8
- Linked courses3
How to use this roadmap
Work through the stages in order. Core steps are the ones everyone on this path needs; optional steps add depth when you have time, and alternatives are other ways to the same skill, so pick one. Steps with a course link to a free course from LearnerMap Academy.
Members of LearnerMap Academy can mark each step as learning, done or skipped, and see how far along the path they are.
Foundations
Python for AI work
CoreWrite clear Python and use arrays for numeric work.
Working with data
CoreLoad, clean and explore tabular data before it reaches a model.
Machine learning concepts
CoreKnow training, validation, overfitting and common metrics, even if you mostly call hosted models.
scikit-learn getting startedMachine Learning Crash Course (Google)
Notebooks
OptionalExplore ideas interactively, then move what works into tested code.
Language models
How language models work
CoreUnderstand tokens, context windows, sampling and why outputs vary between runs.
Prompt design
CoreWrite clear instructions with examples and structure, then iterate against test cases.
Choosing a model
CoreBalance quality, latency, cost and data handling when picking a model for a task.
Open-weight models
AlternativeRun models on your own infrastructure when data or cost requires it.
Building applications
Calling model APIs
CoreHandle authentication, timeouts, retries and rate limits around model calls.
Course: API DesignAnthropic API getting startedOpenAI API reference
Tool use and structured output
CoreLet a model call your functions and return data in a shape your code can trust.
Streaming responses
OptionalShow output as it is generated so long answers feel responsive.
Model Context Protocol
OptionalConnect assistants to tools and data through a shared protocol.
Retrieval
Embeddings
CoreTurn text into vectors so similar meanings sit close together.
Vector search
CoreStore embeddings and find nearest neighbours quickly at scale.
Retrieval-augmented generation
CoreGround answers in your own documents and show where each claim came from.
Keyword and hybrid search
OptionalCombine full-text search with vectors for names, codes and rare terms.
Quality and safety
Evaluation
CoreBuild a test set, score outputs automatically, and track quality across changes.
Securing LLM applications
CoreDefend against prompt injection, data leakage and over-permissive tools.
Responsible AI
OptionalIdentify and manage risks to people affected by an AI system.
Fine-tuning
AlternativeAdapt a model to a narrow task when prompting and retrieval fall short.
Production
Serving and containers
CoreWrap model features in a small service and ship it as a container.
Monitoring latency and cost
CoreTrace each model call with its tokens, latency and outcome.
Caching and rate limits
OptionalReuse answers where safe and back off gracefully when a provider says slow down.
Learning with LearnerMap Academy
LearnerMap Academy runs its learning on LearnerMap, a free platform for organisations. Every course here is free for members of LearnerMap Academy: sign in with the email address LearnerMap Academy knows you by, or with a passkey, and the course opens in your own plan.
Your plan holds at most 3 courses in progress at a time, so you finish what you start before pulling in the next one. Lessons open in order, your progress is saved as you complete each one, and every course ends with a quiz; passing it adds a certificate to your record.
Not a member yet? Ask LearnerMap Academy to invite you. Lesson content, quizzes and everyone’s progress stay private to members; these public pages show only what LearnerMap Academy chose to publish.