This is the roadmap we'd give someone starting from zero who wants to become genuinely competent at building with Generative AI — not just prompting ChatGPT well, but shipping real LLM-powered applications. It's the same sequence our Generative AI course compresses into 42 hours; use it as a self-study path, or as a map of what the course covers and why it's ordered this way.
Time estimates assume a few focused hours per week of self-study. Everything below is covered hands-on, in this order, across 14 live sessions in our course — with a project shipped for each major stage.
0. Prerequisites (1–2 weeks)
Prerequisites
1–2 weeks- Basic Python: variables, functions, loops, working with lists and dictionaries
- Command line basics: navigating folders, running a script
- Git basics: clone, commit, push — you'll want this for your portfolio projects
- Sign up for a free-tier API key (Groq or Gemini both offer one) so you can start calling a real model immediately
1. AI & GenAI foundations (1 week)
Foundations
1 week- AI vs ML vs Deep Learning vs Generative AI — know the difference precisely
- What an LLM is, how it generates text, and why it hallucinates
- Tokens, context windows, and the model landscape (GPT, Claude, Gemini, Llama, Mistral)
2. Prompt engineering (1 week)
Prompt engineering
1 week- Zero-shot, one-shot, and few-shot prompting
- Role & context prompting, structured JSON outputs
- Build 3–5 prompts for a real task and test them against varied inputs, not just one example
3. Calling LLMs & building a backend (1–2 weeks)
APIs & backend
1–2 weeks- Call an LLM API from a plain Python script — no framework yet
- Understand system vs user messages, temperature, and token cost
- Wrap it in a simple FastAPI backend with streaming responses
- Ship it: a working script or API endpoint that answers questions using an LLM
4. RAG & vector databases (2 weeks)
RAG & vector databases
2 weeks- Understand embeddings and semantic search
- Pick one vector database (Chroma is the easiest starting point) and learn it properly
- Build a full pipeline: chunk a document, embed it, retrieve relevant chunks, generate a grounded answer
5. AI agents & tool calling (1–2 weeks)
AI agents
1–2 weeks- Understand function/tool calling and the agent loop
- Build a simple agent with one or two tools (a calculator, a search API)
- Learn why agents fail — wrong tool selection, infinite loops — before you need frameworks to hide it from you
6. Multimodal AI (1 week)
Multimodal AI
1 week- Image generation and vision models — what they're good and bad at
- Speech-to-text and text-to-speech basics
- OCR and document intelligence for scanned/image-based documents
7. LLM evaluation (1 week)
LLM evaluation
1 week- Build a small test set of representative inputs and edge cases
- Learn programmatic checks vs LLM-as-judge evaluation
- Evaluate retrieval quality and generation faithfulness separately for any RAG system
8. Production AI (1–2 weeks)
Production AI
1–2 weeks- Architecture: backend, LLM calls, and data layer separated properly
- Caching, rate limiting, authentication, and cost tracking
- Logging, prompt versioning, and basic prompt-injection defence
9. Portfolio & job search (ongoing)
Portfolio & job search
ongoing- Turn 2–3 of your projects into portfolio case studies: problem → architecture → tech → outcome
- Publish your code on GitHub with a clear README
- Practice explaining your projects out loud — this is what interviews actually test
If self-study feels slower than you'd like, this entire roadmap — with a shipped project at every major stage and instructor feedback along the way — is exactly what the Generative AI & AI Application Development course delivers in 42 hours. See the full curriculum for the session-by-session version of this same path.
Keep learning: Ready to skip the self-study grind? See the Generative AI course that teaches this exact path, live.