What this roadmap covers
Generative AI moves quickly, so a list of today's tools goes out of date within months. The concepts last much longer: tokens, context, embeddings, retrieval, tool calls and evaluation. This roadmap teaches those concepts in the order they depend on one another, and treats specific frameworks as replaceable.
It is written for people building applications on top of models. Research careers that invent new models need much deeper mathematics, which is a different path.
The generative AI roadmap, stage by stage
Finish each stage's project before moving on. Each one reuses the last.
Stage 1: Python and APIs
Functions, JSON, virtual environments, async basics, and calling a REST API with error handling. Proof: a script that calls a public API, handles failures and saves clean results.
Stage 2: How LLMs work
Tokens, context windows, temperature, embeddings, and why models hallucinate, at the level of a clear explanation rather than the mathematics. Proof: explain to a friend why a model can sound confident and still be wrong.
Stage 3: Prompting and structured output
Clear instructions, examples, system prompts and JSON schemas so that code can rely on the output. Proof: a tool that turns messy text, such as job posts, into validated structured data.
Stage 4: RAG
Chunking, embeddings, vector search, reranking and citing sources, so that a model can answer from your documents. Proof: a question-answering app over a real PDF collection that shows where each answer came from.
Stage 5: AI agents and tool use
Tool calling, the agent loop, memory, limits and human approval for risky actions. Proof: a small agent that uses two tools to finish a task, stops safely and logs each step.
Stage 6: Evaluation
Test sets, groundedness and accuracy checks, and comparing prompts or models by measurement instead of impression. Proof: an evaluation report on your RAG app, with one change that measurably improved it.
Stage 7: Deployment and cost
FastAPI, Docker, streaming, caching, logging, rate limits and protection against prompt injection. Proof: your app live at a public URL with usage and cost logged.
The tools, by job rather than by brand
Learn one option in each row. The concepts transfer when tools change.
Model access
A hosted model API from a major provider, and an open model through Hugging Face or Ollama, so you understand both approaches.
Retrieval
An embedding model and a vector store such as FAISS, Chroma or pgvector.
Orchestration
Plain Python first. Add LangChain, LangGraph or LlamaIndex when you can say what problem they are solving for you.
Serving
FastAPI and Docker, with a simple Streamlit or web front end for demos.
Common mistakes on this path
Most stalled projects come from one of these.
- Starting with an agent framework before you can make one reliable model call.
- Building RAG without checking whether the right passages are retrieved.
- Judging quality by trying three questions instead of keeping a test set.
- Fine-tuning a model when better retrieval or a clearer prompt would solve the problem.
- Ignoring cost and latency until the demo is finished.
Where to start from your background
The stages are the same, but your first few weeks differ.
You are new to coding
Spend real time on Stage 1. Weak Python becomes the bottleneck in every later stage.
You are a software developer
Move quickly to Stage 3. Your advantage is production thinking, so go deep on evaluation and deployment.
You work in data or ML
Focus on retrieval and agents, and bring your testing discipline to evaluation.
Quick answers about the generative AI roadmap
Short answers to the questions beginners ask most.
Do I need machine learning before generative AI?
Not to build applications. You need solid Python and a working idea of how LLMs behave. Classic ML helps later, especially for evaluation and ranking.
How much maths does generative AI need?
Very little for application work beyond understanding similarity and probability at a basic level. Research roles need far more.
What is the most important skill to learn?
Evaluation. Many people can build a demo, but fewer can show that it works reliably and improve it with evidence.
Which framework should I learn first?
None at first. Build a small RAG app in plain Python, then a framework will make sense as a time-saver instead of a mystery.
What project should a beginner show?
A RAG app or small agent on a real problem, deployed publicly, with an honest evaluation in the README.
Go deeper on each stage
Guides and programme pages that fill in this roadmap.
Make one model call today
Write a short Python script that sends a real piece of text to a model and gets validated JSON back. Add one failure case and handle it. Everything else in this roadmap grows from that small loop.

