A five-step path from "I can code" to shipping applications built on language models — hosted APIs,
open-source alternatives, retrieval over your own data, and enough of the internals to debug what
the API hides. This is the longest path on the site, and the one that ends with something you can deploy.
Who this is for: you're comfortable writing code in Python or JavaScript and want to
build applications on top of AI models — chatbots grounded in your own documents, agents that call
tools, features powered by an LLM — rather than just use AI products someone else built.
Who this is not for: if you don't code yet, start with the
AI Fundamentals path —
or the No-Code AI Automation path
if you want to build AI-powered workflows without writing code. If your app architecture is fine but the
model's outputs are unreliable, the shorter
Prompt Engineering path
may be the actual fix.
The sequence
1
Survey the whole GenAI application stack once
Goal: build a working vocabulary for every piece of an AI app — models, prompts, embeddings, RAG, agents, fine-tuning — before going deep on any single one.
Microsoft's 21-lesson course is deliberately breadth-first, with runnable code samples in both
Python and TypeScript for nearly every lesson. If you finished the AI Fundamentals path you've
already met this resource — this time, code along instead of skimming, and skip only the lessons
covering ground you genuinely know.
Microsoft — Generative AI for BeginnersWebsite · Hands-On Coding · Beginner
Ship your first real features against a hosted API
Goal: get comfortable calling a frontier model from your own code — chat, structured outputs, function calling, embeddings — the four patterns nearly every AI feature is built from.
The OpenAI Cookbook is copy-paste-ready recipes for exactly these patterns, and most production AI
features today are some combination of them. Don't just read: pick three recipes and get each one
running in your own environment with your own inputs before moving on.
Goal: know when to reach for an open model you run yourself versus a hosted API — and be able to actually do it with the standard tooling.
Hugging Face's LLM Course covers using, evaluating, and fine-tuning open models with the same
libraries used in production everywhere. Even if you end up shipping on hosted APIs, knowing the
open-source path keeps you from being locked into one vendor's pricing and capabilities.
Hugging Face (courses)Website · Hands-On Coding · Intermediate
Build the canonical AI app: retrieval over your own data
Goal: understand chunking, embeddings, retrieval, and generation well enough to build a document chatbot that cites real sources instead of inventing them.
Retrieval-augmented generation is the single most-requested AI app pattern, and this channel's RAG
playlist is the deepest free code-along coverage of it — chunking strategies, embedding choices,
retrieval quality, and the failure modes you'll actually hit. Watch with an editor open, not passively.
Go under the hood so you can debug what the API hides
Goal: build a small GPT from scratch once, so tokenization, context windows, and sampling stop being magic when your app misbehaves.
Karpathy's "Neural Networks: Zero to Hero" series builds language models from first principles in
plain code. You can ship AI apps without this step — but developers who've done it debug faster,
estimate feasibility better, and stop cargo-culting other people's parameter settings.
Andrej KarpathyYouTube · Hands-On Coding · Intermediate-Advanced
Build and deploy a RAG chatbot over documents you actually care about
Pick a real corpus — your personal notes, a hobby wiki, your team's internal docs — and build a
chatbot that answers questions grounded in those documents, with citations back to the source
passages. Add at least one function-calling tool (a lookup, a calculator, a search). Then deploy it
somewhere a friend can actually use it, even if that's just a tunnel to your laptop.
You're done when you can show:
Answers that cite the actual source passages they came from — and say "I don't know" when the corpus doesn't cover it
At least one question that triggers a real tool call instead of a text-only answer
A short write-up of one hallucination or retrieval failure you caught, and what you changed to fix it
Where to go next
If step 3 was the most interesting part, Fast.ai on the dashboard goes much deeper
into training and fine-tuning models yourself, and the Research & Theory category
builds the reading skills to follow new papers. If you want structured, graded curricula with
certificates, DeepLearning.AI and Coursera on the dashboard cover
that. And if your app works but the model's behavior is inconsistent, the
Prompt Engineering path
is a focused two-week fix you can run in parallel.
This path was last reviewed and confirmed against live resource listings on 2026-07-22.