Become an AI Engineer in 2025: The 6-Step Roadmap

Career Compass
4 followers
Oct 26, 2025
Want to start a career in Artificial Intelligence but not sure where to begin? This video breaks down a clear, practical 6-step roadmap to becoming an AI Engineer in 2025 — from mastering Python and math fundamentals to building real-world AI projects and landing your first job.
If you haven't noticed, AI and language models are changing the skills needed to be a successful
engineer.
The people who are learning these new skills are getting jobs that pay up to $435,000 a
year.
They're creating apps using language models that generate millions of dollars at 90% margins
with no employees.
They're building features that took months in minutes.
And they're using AI engineering.
But there's a problem.
These new AI models are non-deterministic, and English is now a programming language.
What?
And frankly, the industry is still figuring out the best practices of using them.
However, we're now entering a golden era for AI engineers, and the ones that add to their
skill sets will be ahead of the curve.
There are six new AI engineering skills that are actually solving these problems, and in
my opinion, every engineer is going to need to learn them to be successful.
For those that are new to the channel, my name is Greg.
I led a team of data scientists at Salesforce, driving product analytics for billion-dollar
products.
I've taught over 10,000 developers how to build their first AI applications, and now
I'm in the absolutely privileged position to be making money doing what I love, building
and teaching with others.
This video is all about telling myself what I wish I would have known about working with
LLMs, and some of these areas I completely overlooked.
I'm going to outline the roadmap of skills that you need to be a successful AI engineer.
Along the way, I'm going to highlight specific jobs that these skills are required for, I'll
show you which experts are demonstrating these skills at a professional level, and I'll give
you a list of resources to help you go further.
And you invest in these skills, you're going to go off and build way more than you ever
thought that you could.
If you're remotely serious about getting into AI engineering, then you're going to want
to watch the full thing.
And by the way, every link and mention in this video is going to be consolidated into
a one-pager.
If you want to get access to it, link's in the description.
The first stop on the roadmap of AI engineering is working with models.
Now there's four popular models.
This is going to be OpenAI, Anthropic, Meta, and Google, and each one has their own personality.
Check out how Arjun Kannan, CEO of ResiDesk, puts it.
We found that the OpenAI models tend to be the best analysts, the Anthropic models tend
to be the best writers, and the Gemini models broadly tend to be the best detectives, if
you want to think about it that way.
The important part to understand is that most usage is going to be from text to text.
However, there is text-to-speech, text-to-image, video-to-image, really any modality that you
may want.
And this is the multi-modal part of working with these models.
The next piece you're going to want to understand is how to work with these models' APIs.
Now you're going to get very used to seeing OpenAI.chatCompletions.create.
But then there's also a lot of different model types, GPT-40, O1, Whisper, Sonnet, HiQ, Opus.
There's really a ton to learn.
You're also going to want to understand streaming, batch processing, prompt caching, and assistance.
Now for each one of these on the document that I talked about beforehand, we're going
to have definitions and resources to go learn more about these.
Next up is going to be local versus open source.
Now if you don't trust OpenAI, Anthropic, Google, or Meta, well then you can go run
your own model on your own infrastructure.
And popular examples for these are going to be with OpenRouter or Ollama.
Now this is up to you.
Now in my opinion, what you're not going to have to worry about yet is with regards to
model training, fine-tuning, or model routing.
Now all these are very important, but I don't think that they're mission critical for an
AI engineer right when they're starting.
It's more of an optimization skill that they can learn later on down the road.
For resources, if you want to learn more, check out my previous AI Show & Tell videos
and we have different guests that talk about their experiences with working with models.
Now another one I would highly suggest is the Lex Friedman Podcast with the Cursor Team.
This is a bit more advanced look into model management, but I thought it was really, really
good.
And for people who to follow, I think Justine Tunney is absolutely fabulous for this.
Now heads up again, once more, this video is going to have a ton of links and resources.
It's too many to put in the description of this video, so I have a doc for one of those.
And if you want that doc, link is in the description for those.
Next up on the roadmap to becoming an AI engineer is understanding the art of prompting.
Now I know a lot of people will argue with me and say prompting isn't serious or that
it is a fad, but the way that I see it is with prompting, what you're really doing is
just trying to elicit a behavior from a model.
Now today that's English, tomorrow it might be something else.
So let's not argue about the future, but let's agree today that effective prompting or the
art of eliciting the behavior from a model that you want is a skill that's needed.
Now the art of prompting goes extremely deep as a practice.
So we won't go into all the details in this video, but I'm going to share four techniques
that have worked for me in the past.
Now the first one that I love is chain of thought or think out loud.
This is going to be when you have the model explain its thought process first, rather
than have the model derive its answer early in its output.
Next it's hard to get performance from a prompt if you don't include examples.
So make sure you include those.
And then of course there's a slew of other tricks like using XML tags or using pre-filled
prompt messages.
If you want to find out what those are, those are on the doc as well.
The next technique you're going to want to master is with structured outputs.
Now it's interesting because computers do much better with structured data like JSON
or tables.
However, language models, they deal with plain text.
Literally, the output is a text that you have, the normal output.
Now when the model responds back to you with valid JSON, well, then you can go and give
that to other programs in your computers, or you can start to use other tools outside
of your language models.
And if you want your language models to use things like search queries or tools or integrate
with other systems, well, structured outputs are going to be the way that you're going
to do this.
Now let's listen to what Thibaut says, and he sold his business for $8 million.
Now he's heavily involved with AI.
I think the sentence that I type the most is, output in JSON, do not output anything
else.
So it's basically structured outputs only that come out from it.
Yeah.
And super reliable.
I'll put a link to how to deal with structured outputs on the doc.
Next up, as an AI engineer, you're going to come up to the question around how to do prompt
management.
And one of my favorite people to follow in this case is Jared, the CEO of PromptLayer.
He has a super cool perspective on this.
It's interesting to see that we kind of all go down the same path.
First we're going to put our prompts right in our code.
Then we're going to have .txt files.
Then we're probably going to use an outside manager, prompt manager, like a PromptLayer.
For resources, we have a ton of these.
First, I want you to check out this Anthropic job, and it's not current anymore, but it
was there, that paid up to $375,000 a year for a prompt engineer or librarian.
Hmm, a prompt librarian, okay.
Next up, we're also going to have a well-rounded introduction to optimize prompting.
And I highly suggest reading Eugene's post.
And for a huge deep dive, well, Elvis is the master in this one.
He is the prompt engineering guy that I highly suggest going and checking out.
And not only that, but we also have research directions as well.
So if you want to check out this paper that Google just came out with around prompting,
links in the doc as well.
The third skill on the AI engineering roadmap is around context or retrieval.
So far, we've only been working with the model as is.
That is, we've relied on the data that they model new at its time of training.
And this is okay for general reasoning, but what about when you want to combine it with
the context it knows about you, or your users, or maybe your role?
The popular term for this is retrieval, as in, hey, we're going to go retrieve data and
bring it back to our language model so it can generate a better response.
This is where the term retrieval augmented generation, or RAG, comes in, because you're
doing generation with retrieval.
But don't be intimidated because this is literally just putting data in your prompt like you
would for any other analysis.
I do this all the time when I copy and paste information about myself into chat GPT.
That's a really manual form of retrieval, but I'm still retrieving information about
myself.
The common way to do this is with matching relevant documents based off of a user's query.
And what's amazing is that the most popular way to do this has been with embeddings and
semantic search.
Let's break both these down very quickly because they're super important.
So embeddings are a vector representation of a body of text.
If that sounds complicated, it's actually pretty simple.
This is literally just a list of numbers, a vector, and we do this because computers
can compare numbers a lot more easily than they can compare words.
So if we have two vectors together, which represents two different pieces of text, we
can actually compare those and do similarity metrics between the two.
And then all of a sudden we can tell, hey, how semantically close are these two different
pieces of text?
Now, the other keyword here is semantic or semantic search.
And this means that you're going to be searching by the meaning of a word or phrase rather
than the exact keyword.
So all of a sudden, ocean is semantically similar to water, which is very cool, right?
And you'll see that there are immediate problems with doing vanilla retrieval.
You'll find that users' queries aren't really detailed enough, or your app isn't returning
the exact right context that you need to answer a question, or perhaps maybe there's too much
fluff in your prompts.
And this is where the world of advanced retrieval comes in.
And there's techniques like enhancing users' query before it goes into your search, or
improving the raw data chunking strategies, or improving how you split your long text
into smaller pieces, which is really your index strategies.
For resources, I have an entire series on retrieval called FullStackRetrieval.com.
If you're interested to learn more, you can go check that out.
There's also the LangChain documentation, which is a great place to get ramped up to
concepts.
Also, for those that are ready, I've made a tool called ChunkViz.com that actually helps
you visualize different chunking techniques.
The next AI engineering skill that you're going to want to master is orchestration.
This is when you go just beyond a single API call to a model, and you start creating systems
that work together in concert.
Now the base levels of this look like working with an orchestration framework, something
like LangChain, that's going to help you adopt the foundation orchestration patterns.
Now at its simplest form, this could just be chains, right?
Basic chains are just when you sequentially put different model calls together.
However, the advanced version of this is when you start to dip into the world of agents.
Now agents may seem complicated, and we're still arguing about what the definition of
an exact agent is.
However, my definition is going to be similar to what Harrison Chase says.
The core idea of agents is using the language model as a reasoning engine.
So using it to determine kind of like what to do and how to interact with the outside
world.
And this means that there is a non-deterministic kind of like sequence of actions that'll be
taken depending on the user input.
In English, agents are simply language models that have access to tools and can decide when
a job is specifically done.
But I will say that the line is blurred for what is an agent, what is not.
Let's not argue about it.
There are jobs that will pay up to $435,000 a year for people who specialize in agents.
Now in the world of agents, the most popular frameworks are going to be LangGraph, CrewAI,
and Haystack.
Or in the no-code world, you have things like Lindy.ai or LangFlow or Howie.ai.
Another key important part with working with agents is going to be around long-term memory.
Now LangChain has a really cool blog post on this entire thing.
However, I will say we're all still trying to figure it out.
And I just posted a tweet that asked everybody what they thought about long-term memory,
and there's no solidified answers.
For other resources, I suggest checking out LangChain just put out a state of agents report.
Now this is around agent adoption within companies.
So if you're looking to build a company, it would be pretty helpful to know what the adoption
patterns currently are.
This one's very, very cool.
The people I suggest to follow in this world are going to be Harrison and Alex.
And by the way, if you want to learn more about building simple agents on top of unstructured
data, this is where my cohort comes in.
Next up on the AI engineering roadmap, and this is arguably most important of all, is
evaluations and observability.
One of my favorite lines about evaluations comes from Jason.
If you don't have evals, then you don't have a serious app.
Evals are the unit tests of your language model applications.
Now in regular code, it's easy to see if things break, but what about when your output is
non-deterministic and really vibe based like language models are?
Well, this is where evals comes in.
People will stay away from evals, well, because look how people stay away from regular unit
testing in their code anyway, but really also because it's hard.
It's hard when you want to thoroughly evaluate how good a summary is or how good a natural
language output is.
These things are difficult, but there's evolving best practices.
One of the best resources I know about this is going to be Hamill.
He's one of the leading voices on evals and has a great resource to go learn more.
Now on the observability side, this really breaks down into two different pieces.
The first one is tracing.
Now tracing is when you literally just start to log the different language model calls
you have so you can easily debug for them.
Now it's extremely difficult to dig into why your application performance has degraded
if you can't find or see which LLM calls you're doing.
The best way I've seen to do tracing is just hook into an additional framework, something
like Langsmith or another observability platform.
Make sure to add a bunch of metadata to your calls so you can easily pinpoint and deduplicate
which calls you're actually making.
On the other side, you're going to want to do cost management.
Now if you're not tracking your costs, then you're going to be in for a bad time.
Observability tools help you track every single LLM call and the latency errors as well.
For resources, my go-to observability tool is Langsmith.
They have a free tier which is super easy to get set up with and it's a really easy
interface.
For other tools, check out Gentrace and Arise.
For product teams, I suggest checking out Autoblox and Freeplay.
The final skill that is needed to be an AI engineer is going to be the mindset.
Now I bluffed a little bit because this is really a meta skill, but it's really important
to have a new mindset as you're starting to build with AI.
Now hey, LLMs give us a new capability to work with and when you have new capabilities,
you have new use cases.
What are these things even good for?
Check out this sample of use cases that Elvis talked about.
Sam Parr of Hampton, he put out an AI report about how all his companies are using AI,
which is super interesting.
And then I also put together an early signals.
The next mindset pillar I want you to adopt is the build first and build quickly mindset.
So check out what Thibaut says about building quick.
I would try to get the first idea that I have and start working on it right away and try
to ship something to the public as soon as possible.
And I would tweet about that.
Even if it's like, even if it's shitty and even if it's not that useful, I would just
tweet about it and expect a few of my friends to test it and give me feedback.
The friction is so low that execution is more important than ever.
The next skill is around understanding the emerging new tool stack.
Now I'm sure you've heard of a few of these and they could take up an entire video each.
But if you're building front ends, it's too cheap not to work with something like V0 to
get inspiration from.
Then there's Cursor and Windsurf for if you're looking for a new AI powered IDE.
And then not only that, Anthropx Projects is very popular amongst developers.
See what Nicole Headley, CEO of Headstart, has to say about Anthropx Projects.
We're just top users of Cloud and Cloud Projects.
We love it.
I think it's the greatest thing that has ever been developed.
If you took it away from me, I would really struggle to run my business, which is probably
like the highest NPS you can get for a product.
The next mindset pillar is around scaling LLM apps.
I got to attend OpenAI's Dev Day and I heard an amazing talk by Colin and Jeff of OpenAI.
In it, they highlight three key areas of scaling LLM apps.
Improving performance, reducing cost, and reducing latency.
And by the way, I packaged up all these links and resources in a single blog post completely
free.
Link for that is in the description if you want it.
Now if you've gotten this far in the video, there's a good chance that you're on your
way to becoming an AI engineer.
And I want to keep pointing you in the right direction and this is only the beginning.
The fact is, is that the landscape is changing so much, which really that should encourage
you because there's never been a better time to build or be a builder who's actually executing.
Now I'm going to be stepping a few students through these best practices hands-on through
a course called Building with AIMs, AI Engineering Patterns for Scrappy Developers.
The whole point of this is I want you to ship an app using these exact six skills.
You're going to learn new tools along the way and you're going to be surrounded by a
cohort of builders.
We're also going to have guest speakers from industry who are going to come in and chat
with us about this.
Not only that, but then we're also going to have AI credits.
So you're going to get a speed boost at building with some of these tools in general.
The cohort will be running ongoing.
If you want to get more information about that, link is in the description and the doc
as well.
