Research-backed roadmap · 10 interactive demos · full explanations for all 17 topics

Become an
AI Engineer

A structured roadmap from zero to production AI systems. Every concept has an interactive demo — try it first, then read the numbered explanation below.

6 phases·17 topics·10 demos·85 key insights
AI Engineer (this roadmap)
  • ·Builds with LLMs and AI APIs
  • ·Prompt engineering, RAG, agents
  • ·Ships products quickly
  • ·Entry salary: ~$103k → $121k+
  • ·LinkedIn #1 fastest-growing role 2025
ML Engineer (different path)
  • ·Trains and fine-tunes models
  • ·PyTorch, data pipelines, GPUs
  • ·Deep math & statistics focus
  • ·Entry salary: ~$128k → $142k+
  • ·Steeper barrier to entry
01

Foundations

6–8 weeks

The bedrock everything else depends on.

Python for AI

2–3 weeks

NumPy, Pandas, Matplotlib — the daily tools of every AI engineer.

Why this matters: You'll write Python every single day. Everything else builds on this.
freeCodeCamp — Python for Data Analysis (YouTube)Real Python — NumPy & Pandas Tutorialsfast.ai — Practical Deep Learning Part 1

Math: Vectors, Matrices & Calculus

▶ demo3–4 weeks

Linear algebra gives you intuition for how models transform data. Calculus explains how they learn.

Why this matters: Without this, you can use AI tools but you can't debug, improve, or design them.
3Blue1Brown — Essence of Linear Algebra (YouTube)3Blue1Brown — Essence of Calculus (YouTube)DeepLearning.AI — Math for ML Specialization

Git & Software Engineering Basics

1 week

Version control, virtual environments, command-line fluency.

Why this matters: You need to collaborate on code, track experiments, and ship things.
GitHub Learning Lab — Introduction to GitHubfreeCodeCamp — Git & GitHub Tutorial (YouTube)
02

Core ML

4–6 weeks

Learn how models learn.

Supervised Learning

▶ demo2–3 weeks

Regression, classification, cross-validation, train/test splits, evaluation metrics.

Why this matters: This is the conceptual foundation of 95% of deployed ML models.
Andrew Ng — Machine Learning Specialization (Coursera) ★fast.ai — Practical Deep Learning for Coders (free)

Neural Networks from Scratch

▶ demo2–3 weeks

Perceptrons, forward pass, backpropagation, activation functions, gradient descent.

Why this matters: You cannot debug or improve models you don't understand at a mechanical level.
Andrej Karpathy — Neural Networks: Zero to Hero (YouTube) ★3Blue1Brown — Neural Networks Series (YouTube)Michael Nielsen — Neural Networks & Deep Learning (free book)
03

Deep Learning

4–6 weeks

The architecture that powers modern AI.

CNNs & Transfer Learning

2 weeks

Convolutional filters, pooling, pre-trained models (ResNet, EfficientNet).

Why this matters: CNNs process images, medical scans, video — and transfer learning is how you adapt them cheaply.
fast.ai — Part 1, Lessons 1–7 (free)CS231n — Convolutional Neural Networks (Stanford, free)

Transformers & Self-Attention

▶ demo2–3 weeks

The architecture behind GPT, Claude, Gemini — every modern LLM.

Why this matters: Every AI product you build will use transformers. This is the single most important architecture of the decade.
Jay Alammar — The Illustrated Transformer (blog, free)Andrej Karpathy — Let's build GPT from scratch (YouTube)DeepLearning.AI — Generative AI with LLMs (free audit)
04

LLMs & Modern AI

4–6 weeks

From model weights to products people use.

Tokenization

▶ demo2–3 days

LLMs don't read words — they read tokens. BPE, SentencePiece, context windows.

Why this matters: You're billed per token. Context windows are in tokens. Wrong tokenization = bugs in prompts.
Andrej Karpathy — Let's build the GPT Tokenizer (YouTube)Tiktokenizer — interactive tokenizer playground

Sampling & Temperature

▶ demo1 week

Temperature, top-k, top-p sampling — the controls that shape LLM output.

Why this matters: Prompt engineering without understanding sampling is guesswork. This gives you control.
Anthropic — Claude Prompt Engineering Guide (free)DeepLearning.AI — ChatGPT Prompt Engineering for Developers (free)

Hallucinations & Reliability

▶ demo1 week

Why LLMs confidently generate falsehoods, and how to prevent it.

Why this matters: This is the #1 failure mode in production AI. You must understand it to build reliable systems.
Anthropic — Constitutional AI & AI Safety Research (free)Braintrust — Building Reliable LLM Pipelines (blog)

Embeddings & Semantic Search

▶ demo1 week

How words become vectors. Cosine similarity, semantic clusters, embedding models.

Why this matters: Embeddings are the foundation of RAG, vector DBs, and semantic search — all of Phase 5.
Jay Alammar — The Illustrated Word2Vec (blog, free)OpenAI — Embeddings API Guide (free)Sentence Transformers — Quickstart (free)
05

Production AI Systems

6–8 weeks

Ship AI that works in the real world.

RAG — Retrieval Augmented Generation

▶ demo2–3 weeks

Ground LLMs in real documents. Chunking, vector retrieval, prompt injection.

Why this matters: The most common AI engineer task. RAG is how you give LLMs up-to-date, factual, private knowledge.
DeepLearning.AI — Retrieval Augmented Generation (free)LlamaIndex — Documentation & Tutorials (free)Activeloop — Advanced RAG Course (free audit)

AI Agents & Agentic Workflows

▶ demo3–4 weeks

Observe → Plan → Act → Reflect loops. Tool use, multi-agent systems, memory.

Why this matters: Agents are the frontier. Products that complete multi-step tasks autonomously are the next generation of AI software.
Coursera — Building AI Agents & Agentic Workflows (Specialization)LangGraph — Documentation (free)CrewAI — Documentation (free)DeepLearning.AI — Multi-Agent Systems with Claude (free)

Evaluation & Testing

1–2 weeks

LLM-as-judge, RAGAS, PromptFoo, regression suites. Ship only what you can measure.

Why this matters: You cannot iterate on what you can't measure. Evals are the CI/CD of AI systems.
Braintrust — LLM Evaluation Guide (free)RAGAS — RAG Evaluation Framework (open source)PromptFoo — Prompt Testing CLI (open source)

Deployment & LLMOps

2 weeks

FastAPI, Docker, monitoring, cost optimization, CI/CD for LLM systems.

Why this matters: A demo that doesn't ship is a hobby. LLMOps is how you turn demos into products.
Coursera — MLOps and LLMOps (1 week)DeepLearning.AI — LLMOps Short Course (free)FastAPI — Documentation (free)MLflow — Model Tracking (open source)
06

Portfolio & Projects

Ongoing

Ship things. Get hired.

What to Build

2–8 weeks each

Employers care about shipped products, not certificates. 3 projects by difficulty.

Why this matters: A public GitHub with live demos > 10 certificates. Build in public from day one.
Beginner: AI Chat with PDF (RAG + FastAPI + Vercel)Intermediate: Fact-Checking Agent (CrewAI + web search)Advanced: Multi-Agent Research System (LangGraph + RAG + Evals)

How to Land the Job

Ongoing

Public GitHub + live demo + metrics in the README + blog post = a winning portfolio.

Why this matters: Hiring managers spend 90 seconds on your portfolio. Make those 90 seconds count.
Vercel — Deploy in seconds (free tier)Hugging Face Spaces — Free ML demos hostingdev.to / Hashnode — Write about your projects (free)
You have the roadmap.
Pick Phase 01. Open the first resource. Write the first line of Python. The only way to become an AI engineer is to start.
Foundations·Core ML·Deep Learning·LLMs & Modern AI·Production AI Systems·Portfolio & Projects