For the past months, whenever someone asked me how to get into AI, I had no good single answer. The material out there is either academic theory, marketing hype, or fifty scattered tutorials that each assume you read the other forty-nine.
So I wrote the thing I wanted to hand over: Artificial Intelligence in Practice, an 84-page handbook that starts at zero and ends with you running agents with tools. It is free, and it just went up on GitHub:
https://github.com/vinimabreu/ai-in-practice
Who it is for
People starting out in AI, or curious about it, who want one coherent path instead of a pile of tabs. It assumes no AI background. If you can open a terminal, you can follow it. Developers already building with LLMs will find parts of it review, though the evaluation and cost chapters may still earn their time.
The path it walks
- Foundations. What tokens, context windows and temperature actually are, and why a model "hallucinates". Setting up a real dev environment without pain.
- Running models on your own machine. Ollama, picking a model family for your task, and honestly sizing what your hardware can run, including Apple Silicon.
- RAG. Making an AI answer from YOUR documents instead of making things up, explained without hand-waving: embeddings, vector databases, chunking, and when RAG is the wrong answer.
- Agents. The loop that turns a chatbot into something that does work: CrewAI, LangGraph, tools, browsing, MCP, and agents from different frameworks talking to each other.
- Going deeper. Fine-tuning vs RAG, what the APIs really cost, security and privacy of autonomous agents, prompting as a method, and how to test AI systems so they do not embarrass you in front of a customer.
- Three capstone projects, from a document Q&A assistant to a research agent crew, ending with what it takes to go from project to product.
Everything in it was tested by hand. Prices and version numbers were verified in mid-2026, and the book tells you to re-verify them, because this field does not sit still.
Why free
Because the version of me from a few years ago needed this and could not have paid for it. It is under CC BY-NC-SA: share it, translate it, use it to teach a course. Just keep it non-commercial and credit the source.
Two small asks
If you know someone who keeps saying "I want to learn AI but I do not know where to start", send them the link. That is exactly who it was written for.
And if it is useful to you, a star on the repo helps other people find it: https://github.com/vinimabreu/ai-in-practice
I will be serializing some of the deeper chapters here on dev.to over the next weeks, starting with the one I think is most neglected: how to actually test an LLM system. If there is a chapter you want first, tell me in the comments.
Vinicius Pereira
Top comments (8)
Brilliant, and thank you! How come this doesn't have more likes, and your repo doesn't have more stars? Starred it, and of course planning to read it ...
P.S. skimmed it - I'm impressed by how well-written this looks, beautiful graphics too - professional job!
Thank you, genuinely. The star means more than the number suggests, because this one was a long solo stretch and you are the first person to ask about it out loud.
It came out of real questions. Friends and people around me kept asking the same things, what a token actually is, why the model made something up, whether any of this is usable for real work, and I got tired of writing the same explanation in messages. So the book is shaped by what people actually got stuck on rather than by what a curriculum says should come first.
As for the stars, I think I am simply much better at writing the thing than at telling anyone it exists. Working on it.
If you know someone stepping into this field, pass it along. That is exactly who I had in mind while writing, and it would do more good in their hands than in a star count.
It looks brilliant, very well structured and organized - I certainly intend to spread the word whenever there's an opportunity to do so! (but of course I'm first going to read it myself)
Great stuff, thank you for this!
That is generous, thank you. No rush on the reading either, it is built to be opened at whatever part someone is stuck on rather than read front to back.
If anything in it is wrong or unclear once you get into it, tell me. That kind of note is worth more to me than the spreading.
Yeah if have feedback on the contents I'll let you know!
Hey I'm reading it, it's brilliant - I like the short chapters, which has "pros" and "cons" (but the "cons" are also really "pros" when you think about it, haha):
pros: the short chapters make it a quick and pleasant read! reading it doesn't feel like "homework" or a "chore" - you make a point or you explain something, the reader notes it, and on to the next topic
cons: in many cases the info is very basic, you outline a concept or a technique and then it's up to the user to find out more
Well so actually I think the "cons" are really "pros" in disguise - you just give the reader the essentials, and then it's up to them to further explore it themselves! I think that works better than spoon feeding the reader all the details and overloading them with information (which, in this fast moving field is going to be outdated quickly anyway) ...
I have one tip (or "complaint", haha): the contents is divided into "parts" (Part I till Part X), but in the TOC (Table of Contents) I only see the "chapters", not the "parts" - it would be helpful if the Parts were listed in the TOC ...
One concept I feel is missing at the moment: skills! That's being talked about a lot (even with people asserting that MCP isn't so important anymore, because we have 'agents' and 'skills' - no idea if that makes sense or not) ...
And one limitation I ran into personally - the idea of running the open source models locally is nice, but I quickly realized that I'm lacking the hardware - 8 GB memory is insufficient for anything but the tiniest (quantized) models - not enough to do any sort of serious work, or even experimentation ...
I then looked at the options to run open source models on 'rented' VMs in the cloud (AWS etc), but that wasn't a solution either - way too costly for anything but the tiniest model, again ...
So I thought I hit a dead end there, and the only solution would be to use a paid API (Anthropic or OpenAI), until I came across this - the NVIDIA NIM service:
build.nvidia.com/models
dev.to/fretny/nvidia-nim-is-mind-b...
mindstudio.ai/blog/nvidia-nim-free...
Looks brilliant - as good as free for experimentation (even for low traffic/volume production), with very capable open source models - the only real limitation are the rate limits, but for experimentation and learning that's not really an issue ...
This is the feedback I was asking for, so thank you for actually taking the time.
Table of contents per part: agreed, and it is a real gap. Eighty-four pages without internal signposts asks the reader to hold the map in their head. Going into the next revision.
Skills is the sharpest catch. The book walks from tokens to tools to agents and skips the layer in between: tools answer what an agent can do, skills answer how it knows how to do it well, packaged so it loads them when the task calls for it. That is a chapter rather than a paragraph, and it belongs right after tools.
On local models and hardware, fair hit. I was encouraging where I should have been specific. "Run it locally" means something very different at 8 GB of VRAM than at 48, and quantization moves the answer again, so that section needs a plain table of what actually runs on what instead of enthusiasm.
NVIDIA NIM I have not run myself, so I will test it before putting it in the book. If it does what you describe, it fills exactly the hole between a laptop that cannot hold the model and a commercial API that ends the experiment, which is the honest gap in that chapter.
Thanks! Good to hear that those points made sense, and you're right that NVIDIA NIM seems to fill the gap between those two options ("local" LLMs, and commercial APIs) - purely coincidental that I came across it, but I'm glad I did, because it seems to just fit the bill ...
I'm gonna read the rest of the chapters, and then I'm planning to get "hands on" and try some stuff - probably using NVIDIA NIM, because running models locally isn't really realistic with the hardware I have ...
Great stuff - brilliant!