12-week AI roadmap
What to study each week to become an AI engineer or a forward deployed engineer, with the interview questions for each week.
Week 1: How LLMs work
Each week: learn it, build one small thing, then practice the questions interviewers ask about it. What you build becomes your interview stories.
Foundations
Weeks 1 to 4Week 1How LLMs workThis week
Learn- Tokens, context windows and why they decide cost
- Embeddings in one sentence
- Temperature and why answers vary
- Pricing: input, output and cached tokens
- LLM Course, chapter 1 · Hugging Face
- Tiktokenizer · Playground
- Let’s build the GPT tokenizer · Andrej Karpathy (video)
Read in our study libraryBuildCall an LLM API from a script. Print the tokens used and the cost of each call.
Example to run and changeEdit the code and run it again. It runs in your browser; the first run downloads Python (about 10 MB).
Week 2Prompting and structured output
Learn- System prompts, examples (few-shot) and clear instructions
- Getting reliable JSON back
- Why small prompt changes break things
- Prompt engineering guide · Claude docs
- OpenAI Academy · OpenAI
Read in our study libraryBuildTurn messy job posts into clean JSON (title, skills, years). Test it on 10 real posts.
Example to run and changeEdit the code and run it again. It runs in your browser; the first run downloads Python (about 10 MB).
Week 3Embeddings and search
Learn- Chunking text and why chunk size matters
- Similarity search and vector databases
- Hybrid search: keywords plus embeddings
- Short courses (RAG, embeddings) · DeepLearning.AI
- OpenAI Cookbook · OpenAI
Read in our study libraryBuildSemantic search over your own notes or docs. Compare it with plain keyword search.
Example to run and changeEdit the code and run it again. It runs in your browser; the first run downloads Python (about 10 MB).
Week 4RAG end to end
Learn- Retrieve, add to the prompt, answer with citations
- Why RAG answers go wrong (bad retrieval, stale data)
- Context engineering: what to include and what to leave out
- Effective context engineering for agents · Anthropic
- Short courses · DeepLearning.AI
Read in our study libraryBuildA Q&A bot over a set of documents that cites the passage it used.
Example to run and changeEdit the code and run it again. It runs in your browser; the first run downloads Python (about 10 MB).
Build like a pro
Weeks 5 to 8Week 5Evals: proving it works
Learn- Golden test sets
- LLM-as-judge and its limits
- Catching regressions before users do
- OpenAI Cookbook (evals) · OpenAI
- Short courses (evals) · DeepLearning.AI
Read in our study libraryBuildWrite 30 test questions for your week 4 bot, score it, change one thing, score it again.
Example to run and changeEdit the code and run it again. It runs in your browser; the first run downloads Python (about 10 MB).
Week 6Agents and tool use
Learn- Tool (function) calling
- Workflows vs agents, and when an agent is the wrong choice
- MCP: connecting models to tools and data
- Guardrails and human approval
- Building effective agents · Anthropic
- Model Context Protocol · modelcontextprotocol.io
- Anthropic Academy · Anthropic
Read in our study libraryBuildAn agent with two tools (for example search and a calculator) that asks before any action.
Example to run and changeEdit the code and run it again. It runs in your browser; the first run downloads Python (about 10 MB).
Week 7Frameworks: LangChain and LangGraph
Learn- Graphs, state and memory
- Human-in-the-loop steps
- When to skip a framework and write plain code
- Introduction to LangGraph · LangChain Academy
- LangGraph docs · LangChain
Read in our study libraryBuildRebuild your week 6 agent as a LangGraph graph with an approval step.
Example to run and changeEdit the code and run it again. It runs in your browser; the first run downloads Python (about 10 MB).
Week 8Production: cost, speed, safety
Learn- Prompt caching and batching
- Latency budgets and streaming
- Logging and tracing every call
- Prompt injection and personal data
- Prompt caching · Claude docs
- Batch processing · Claude docs
- Token counting · Claude docs
Read in our study libraryBuildAdd cost per request, caching and a log of every call to your project. Cut its cost and write down by how much.
Example to run and changeEdit the code and run it again. It runs in your browser; the first run downloads Python (about 10 MB).
Get hired
Weeks 9 to 12 · AI engineer trackWeek 9Fine-tuning vs RAG vs prompting
Learn- What fine-tuning changes and what it can’t
- Smaller, cheaper models for narrow jobs
- Choosing with data, not taste
- LLM Course · Hugging Face
- Claude developer docs · Anthropic
Read in our study libraryBuildTry one task three ways (prompt only, RAG, a smaller model). Compare quality, cost and speed in a table.
Week 10System design for AI
Learn- Design an AI support assistant end to end
- Data flow, evals, fallbacks and monitoring
- Trade-offs you must be able to defend
- Building effective agents · Anthropic
Read in our study libraryBuildDraw the design of your project: components, data, evals, costs, what fails and what happens then.
Week 11Portfolio and your stories
Learn- A README that shows the problem, the result and the numbers
- Turning projects into STAR stories
BuildPublish one project with a short demo, your eval results and your cost numbers.
Week 12Interview sprint
Learn- Mock interviews, out loud
- Fix your weakest area
- Questions to ask them
BuildTwo voice mocks this week. Retry every answer that scored under 70.
Free courses worth your time
All free. Pair them with the week you’re on; the study library chapters above cover the interview angle.
- Elements of AI · University of Helsinki
- Python · Kaggle Learn
- Intro to Machine Learning · Kaggle Learn
- Neural networks (videos) · 3Blue1Brown
- LLM Course · Hugging Face
- Learn Prompting · learnprompting.org
- Short courses · DeepLearning.AI
- Intro to Deep Learning · Kaggle Learn
- AI Agents Course · Hugging Face
- LangChain tutorials · LangChain
- MLflow docs (MLOps) · MLflow
- Practical Deep Learning · fast.ai
- CS50 AI with Python · Harvard
- Machine Learning Crash Course · Google
- Deep RL Course · Hugging Face
0 of 12 weeks done. Ticks are saved on this device. Behind? Keep going at your own pace; finishing beats speed. Courses and docs are outside sources, reviewed October 2026. No course or roadmap guarantees a job or a salary. Practice only, never during a live interview.