AI Roadmap for Beginners
Build your first small AI projects and choose the next learning path.
Start with basic Python and small data tasks. Learn the maths needed to understand your first machine-learning model, then practise evaluating its predictions. Try a small generative-AI application and check its outputs. Use these projects to choose between data science, AI application development and model engineering.
Quick answer
What is the right AI roadmap for beginners?
Start with basic Python and small data tasks. Learn the maths needed to understand your first machine-learning model, then practise evaluating its predictions. Try a small generative-AI application and check its outputs. Use these projects to choose between data science, AI application development and model engineering.
Sources and methodology · This roadmap is reviewed when production practices, tools or platform patterns materially change.
Stages
8
Last reviewed
16 September 2026
Stage 1: Understand AI and choose a small problem
AI, ML and generative AI; capabilities, limits and task framing.
Understanding what AI can and cannot do prevents wasted effort on impossible projects.
- What you learn
- AI, ML and generative AI concepts.
- Capabilities and limits.
- Task framing.
- Non-AI alternatives.
- What you should build
- Describe an achievable task and a simple non-AI alternative.
- Ready when
- You can describe an achievable AI task and explain a simpler non-AI alternative.
- Common mistake
- Choosing a project that requires infrastructure or data you do not have.
- Acceptance checks
- Describe an achievable task and a simple non-AI alternative.
- Related resources
- Google ML Crash Course — Foundational ML concepts
Stage 2: Python and basic developer tools
Variables, control flow, functions, files, errors, environments and Git basics.
Every later stage depends on running and debugging Python code.
- What you learn
- Variables, control flow and functions.
- Files, errors and exceptions.
- Virtual environments.
- Git basics.
- What you should build
- Run a CSV-processing script and handle a missing file.
- Ready when
- You can run a script from a documented environment and handle a missing-file error.
- Common mistake
- Skipping environment setup and hitting import errors on every run.
- Acceptance checks
- Run a CSV-processing script and handle a missing file.
- Related resources
- Python tutorial — Control flow, data structures and files
- Pro Git — Repository basics
Stage 3: Work with data and introductory SQL
Tables, missing values, duplicates, joins, summaries and plots.
Data quality determines whether any model or analysis is trustworthy.
- What you learn
- Tables and data types.
- Missing values and duplicates.
- Joins and summaries.
- Basic plots.
- What you should build
- Clean and explain a small dataset without losing rows unexpectedly.
- Ready when
- You can clean a small dataset and record row counts before and after joins.
- Common mistake
- Dropping rows with missing values without checking why they are missing.
- Acceptance checks
- Clean a small dataset and explain row-count changes.
- Related resources
- pandas introductory tutorials — Read, clean and summarize tabular data
- PostgreSQL tutorial — Basic SQL queries and joins
Stage 4: Maths and statistics through examples
Vectors, functions, averages, variation, sampling and probability intuition.
Understanding the maths behind models helps you interpret results and avoid errors.
- What you learn
- Vectors and functions.
- Averages and variation.
- Sampling.
- Probability intuition.
- What you should build
- Explain a small-sample limitation and a vector shape mismatch.
- Ready when
- You can explain why a small sample can mislead and debug a shape mismatch.
- Common mistake
- Memorizing formulas without connecting them to a concrete example.
- Acceptance checks
- Explain a small-sample limitation and a vector shape mismatch.
- Related resources
- NumPy beginner guide — Arrays, indexing and shapes
- Google ML Crash Course — Applied maths for ML
Stage 5: Train and evaluate a simple model
Features, labels, baselines, splits, leakage and appropriate metrics.
A model is only useful if it beats a sensible baseline on valid evaluation data.
- What you learn
- Features and labels.
- Baselines.
- Train/test splits and leakage.
- Regression and classification metrics.
- What you should build
- Compare a small model with a baseline on held-out data.
- Ready when
- You can compare a model with a baseline and show representative mistakes.
- Common mistake
- Reporting accuracy without checking the baseline or class balance.
- Acceptance checks
- Compare a small model with a baseline on held-out data.
- Related resources
- Google ML Crash Course — Regression and classification modules
- scikit-learn common pitfalls — Avoiding data leakage
Stage 6: Try a generative AI application
Prompts, tokens, tools/APIs, unsupported answers and privacy.
Generative AI is widely accessible but its failure modes are not obvious without testing.
- What you learn
- Prompts and tokens.
- APIs and accessible tools.
- Unsupported outputs.
- Privacy and simple evaluation.
- What you should build
- Test a summarizer on ordinary and unsupported inputs.
- Ready when
- You can test a generative AI tool on ordinary, ambiguous and unsupported inputs and record limitations.
- Common mistake
- Trusting model outputs without testing failure cases.
- Acceptance checks
- Test a summarizer on ordinary and unsupported inputs; record limitations.
- Related resources
- Hugging Face LLM Course — Tokenizers and pretrained models
Stage 7: Check results and publish your work
Reproducible files, instructions, errors, evidence and limitations.
A project others cannot reproduce does not demonstrate skills.
- What you learn
- Reproducible files.
- README and instructions.
- Evidence and errors.
- Limitations.
- What you should build
- Package your project so another learner can run it and understand what fails.
- Ready when
- Another learner can run the project and understand what fails.
- Common mistake
- Submitting a notebook without a README or environment instructions.
- Acceptance checks
- Another learner can run the project and understand what fails.
- Related resources
- Python tutorial — Environments and modules
- scikit-learn common pitfalls — Validation and reproducibility
Stage 8: Choose your next learning path
Branch into data science, AI application development or model engineering.
The beginner roadmap opens three alternative paths, not a single pipeline.
- What you learn
- Data science path.
- AI developer path.
- ML engineer path.
- Comparing the paths.
- What you should build
- Choose one route based on the work you enjoyed and unmet prerequisites.
- Ready when
- You can choose one route and identify its next unmet prerequisite.
- Common mistake
- Treating the three paths as sequential rather than alternative branches.
- Acceptance checks
- Choose one route and identify its next unmet prerequisite.
- Related resources
- Data Science roadmap — Analysis and statistics path
- AI Developer roadmap — Application building path
Stage 1: Understand AI and choose a small problem
AI, ML and generative AI; capabilities, limits and task framing.
Understanding what AI can and cannot do prevents wasted effort on impossible projects.
- What you learn
- AI, ML and generative AI concepts.
- Capabilities and limits.
- Task framing.
- Non-AI alternatives.
- What you should build
- Describe an achievable task and a simple non-AI alternative.
- Ready when
- You can describe an achievable AI task and explain a simpler non-AI alternative.
- Common mistake
- Choosing a project that requires infrastructure or data you do not have.
- Acceptance checks
- Describe an achievable task and a simple non-AI alternative.
- Related resources
- Google ML Crash Course — Foundational ML concepts
Stage 2: Python and basic developer tools
Variables, control flow, functions, files, errors, environments and Git basics.
Every later stage depends on running and debugging Python code.
- What you learn
- Variables, control flow and functions.
- Files, errors and exceptions.
- Virtual environments.
- Git basics.
- What you should build
- Run a CSV-processing script and handle a missing file.
- Ready when
- You can run a script from a documented environment and handle a missing-file error.
- Common mistake
- Skipping environment setup and hitting import errors on every run.
- Acceptance checks
- Run a CSV-processing script and handle a missing file.
- Related resources
- Python tutorial — Control flow, data structures and files
- Pro Git — Repository basics
Stage 3: Work with data and introductory SQL
Tables, missing values, duplicates, joins, summaries and plots.
Data quality determines whether any model or analysis is trustworthy.
- What you learn
- Tables and data types.
- Missing values and duplicates.
- Joins and summaries.
- Basic plots.
- What you should build
- Clean and explain a small dataset without losing rows unexpectedly.
- Ready when
- You can clean a small dataset and record row counts before and after joins.
- Common mistake
- Dropping rows with missing values without checking why they are missing.
- Acceptance checks
- Clean a small dataset and explain row-count changes.
- Related resources
- pandas introductory tutorials — Read, clean and summarize tabular data
- PostgreSQL tutorial — Basic SQL queries and joins
Stage 4: Maths and statistics through examples
Vectors, functions, averages, variation, sampling and probability intuition.
Understanding the maths behind models helps you interpret results and avoid errors.
- What you learn
- Vectors and functions.
- Averages and variation.
- Sampling.
- Probability intuition.
- What you should build
- Explain a small-sample limitation and a vector shape mismatch.
- Ready when
- You can explain why a small sample can mislead and debug a shape mismatch.
- Common mistake
- Memorizing formulas without connecting them to a concrete example.
- Acceptance checks
- Explain a small-sample limitation and a vector shape mismatch.
- Related resources
- NumPy beginner guide — Arrays, indexing and shapes
- Google ML Crash Course — Applied maths for ML
Stage 5: Train and evaluate a simple model
Features, labels, baselines, splits, leakage and appropriate metrics.
A model is only useful if it beats a sensible baseline on valid evaluation data.
- What you learn
- Features and labels.
- Baselines.
- Train/test splits and leakage.
- Regression and classification metrics.
- What you should build
- Compare a small model with a baseline on held-out data.
- Ready when
- You can compare a model with a baseline and show representative mistakes.
- Common mistake
- Reporting accuracy without checking the baseline or class balance.
- Acceptance checks
- Compare a small model with a baseline on held-out data.
- Related resources
- Google ML Crash Course — Regression and classification modules
- scikit-learn common pitfalls — Avoiding data leakage
Stage 6: Try a generative AI application
Prompts, tokens, tools/APIs, unsupported answers and privacy.
Generative AI is widely accessible but its failure modes are not obvious without testing.
- What you learn
- Prompts and tokens.
- APIs and accessible tools.
- Unsupported outputs.
- Privacy and simple evaluation.
- What you should build
- Test a summarizer on ordinary and unsupported inputs.
- Ready when
- You can test a generative AI tool on ordinary, ambiguous and unsupported inputs and record limitations.
- Common mistake
- Trusting model outputs without testing failure cases.
- Acceptance checks
- Test a summarizer on ordinary and unsupported inputs; record limitations.
- Related resources
- Hugging Face LLM Course — Tokenizers and pretrained models
Stage 7: Check results and publish your work
Reproducible files, instructions, errors, evidence and limitations.
A project others cannot reproduce does not demonstrate skills.
- What you learn
- Reproducible files.
- README and instructions.
- Evidence and errors.
- Limitations.
- What you should build
- Package your project so another learner can run it and understand what fails.
- Ready when
- Another learner can run the project and understand what fails.
- Common mistake
- Submitting a notebook without a README or environment instructions.
- Acceptance checks
- Another learner can run the project and understand what fails.
- Related resources
- Python tutorial — Environments and modules
- scikit-learn common pitfalls — Validation and reproducibility
Stage 8: Choose your next learning path
Branch into data science, AI application development or model engineering.
The beginner roadmap opens three alternative paths, not a single pipeline.
- What you learn
- Data science path.
- AI developer path.
- ML engineer path.
- Comparing the paths.
- What you should build
- Choose one route based on the work you enjoyed and unmet prerequisites.
- Ready when
- You can choose one route and identify its next unmet prerequisite.
- Common mistake
- Treating the three paths as sequential rather than alternative branches.
- Acceptance checks
- Choose one route and identify its next unmet prerequisite.
- Related resources
- Data Science roadmap — Analysis and statistics path
- AI Developer roadmap — Application building path
From roadmap to production
Build production AI Roadmap for Beginners systems with instructor feedback
You have the framework. The Explore AI Engineering training adds what self-study cannot: live instruction, instructor-reviewed labs, production deployment drills and a capstone that proves you can ship and operate — not just understand.
Fees, schedules and enrolment details are on the course page. No placement, salary or outcome is guaranteed.
Capstone
Build and explain your first AI project
Analyse a small public or clearly labelled synthetic dataset and build a classification baseline. An API summarizer is an optional extension. Submit code, a README, held-out results and limitations.
Training alignment
How this roadmap aligns with SCAI courses
This roadmap is free and self-paced. SCAI's AI Engineering course covers the foundations with live instruction and guided projects. A Data Science course is an alternative if analysis resonates more than engineering.
The course adds what the roadmap cannot: instructor feedback on your first projects, peer review, and structured progression through Python, data and ML fundamentals. If you prefer to work independently, this roadmap gives you the full framework.
What to read next
What to read next
After completing this roadmap, choose your next track based on the work you enjoyed. The Data Science roadmap covers analysis, statistics and experiments. The AI Developer roadmap covers building model-powered applications. The ML Engineer roadmap covers reproducible predictive systems and serving.
Related learning
- Continue to the AI Engineer roadmapWhen you are ready to move from foundations to the full engineering track.
- Branch into the ML Engineer roadmapIf model development and experimentation appeal more than broad engineering.
- Move into the Data Science roadmapIf analysis, statistics and business insight resonate more than engineering.
FAQ
Questions about learning AI
Direct answers for beginners.
Can I start learning AI without coding?
Yes. Explore basic concepts while learning Python. The implementation projects gradually require code.
How much maths do I need at the beginning?
Learn the maths used by each exercise. You do not need to finish an advanced maths curriculum before your first small project.
Do I need a powerful GPU?
Not for basic data analysis and small classical models. Larger local models are optional; hosted services can have limits or charges.
Can I try generative AI before machine learning?
Yes. Try a small application early and build data and evaluation skills alongside it.
How long will this roadmap take?
It depends on prior knowledge and practice time. Use the build and readiness checks to decide when to move on.