Transcription
$354,000 is how much some big tech companies are paying AI engineers in 2026. And if you want to become an AI engineer, this is the video for you.
If you're new here, hi, my name is Sad. I got my masters in AI and computer science from Georgia Tech when I was 20. I've worked at big tech companies like Amazon, and I interview popular tech CEOs to get a real insider scoop on the future of AI.
And when it comes to AI engineering specifically, I've seen so many people waste months chasing the wrong skills and over complicating the process. Some people try to jump straight to the flashy AIM ML stuff before building any real foundation. It's like trying to cook a Michelin star meal like Gordon Ramsay before knowing how to use a knife.
And so in this video, we're going to take a step back. I'll first explain what an AI engineer actually is. Then I'll give you the exact five-level AI engineer pyramid you need. Even if you have no experience whatsoever, it's completely fine. I'm going to take you from an absolute foundations of AI engineering all the way to a top-notch Michelin star level. Plus, I'll give you specific resources, skills, and projects you can do to actually get hired so you can become that $350,000 AI engineer fast.
And with that being said, let's get started. First, let's talk about what an AI engineer actually does. So, AI engineers are software engineers who build applications powered by intelligent models. To be clear, a software engineer might build a website like Amazon.com where customers can place and track orders for various items. Then an AI engineer might come in and add an AI chatbot to that same website. So instead of clicking through menus to track your order, you can just type "where's my package?" and the chatbot understands your question, checks your order status, and gives you a clear answer. AI engineers take software products and integrate powerful AI models so they become more intelligent and user-friendly. Software engineers create the dish. AI engineers add the flavoring and garnish on top. Very important.
AI engineers are software engineers and they are the most powerful when they first master software engineering because you cannot enhance a system you do not understand.
And with that being said, let's dive into layer one of our road map known as the SALT foundation. Like I said before, when you're learning to cook, you don't immediately go and start making elaborate beef Wellington dishes. You're not Gordon Ramsay yet. Instead, you start with the basics. How to properly hold a knife or boil water. And just like that, before we master AI engineering, we need to slow down and learn the basics of software engineering. And this can be broken down into SALT.
First, S, software fluency. This means you need to learn a programming language. And if your goal is AI engineering, then you must learn Python because pretty much any of the popular AI or machine learning libraries exist within Python. Python is the kitchen where AA engineers start cooking. But learning Python does not mean watching a few tutorials. Please do not fall into tutorial hell. You learn nothing. Instead, what I really like using is this platform called Cody.te because here you're given bite-size practical lessons and immediately are followed by exercises and many projects. So, you are forced to apply what you just learned. On top of that, they have an AI assistant built into the platform. So, if you get stuck, you can ask questions in real time. It's like having a professional tutor helping you all the way through. And it's completely free to start. But if you do want premium features like unlimited AI queries, I have a discount code down below in the description.
Next, we got A, which stands for API architecture. At a high level, an API is a structured way for two systems to communicate with each other in software. Think of it like a mailman. If I want to send a letter to one of my friends, I don't just waltz into his house and give him the mail. Instead, I write a message, put it in an envelope, hand it to the mailman, and then the mailman follows a system to deliver it to him. In software, that mailman is the API. If you have a website that needs data from a database, it doesn't directly access it. Instead, it goes through an API, which serves as the middleman to get the data and return it to the front-end website. Google Maps uses APIs when trying to collect location data. Amazon uses it when trying to collect customer data. AI engineers use API endpoints to access AI models.
Then we got L&T. L is for life cycle and version control. For this, make sure you learn Git and GitHub because especially when you're collaborating with people as a software engineer, these come in handy. This is like having utensils in the kitchen. Like you got to have them. That means you got to learn your Git and GitHub. It's just like baseline. T is your tech stack, which are the primary technology tool sets you'll be using as a software AI engineer. Learn databases with MongoDB. Backend frameworks like Flask and Node. And for front end, frameworks like React, Angular, or Vue allow you to build dynamic UIs to connect everything and build a wonderful full stack application.
And at this point, you should be pretty confident in your foundational skills. But we don't want to just learn how to cut a knife or season food with salt. No, we must spice things up with the next level, layer two, known as controlled intelligence. Here, we'll learn how to cook by following specific recipes. This means we take our software engineering skills and level up with basic AI integrations. That's why it's called controlled intelligence.
In a practical sense, for your Python code, here you will learn how to call OpenAI APIs or utilize Hugging Face models. For those who don't know, Hugging Face is like a marketplace for AI models. So instead of training your own model from scratch, which is incredibly difficult and takes a lot of data and computing power, you can just download pre-trained models from Hugging Face. And especially if you're a beginner, you need to do this because now it's not the time for you to be inventing new recipes. Instead, you should learn the art of cooking by following a recipe that already works. And that's exactly what Hugging Face allows you to do with AI. And the beauty of Hugging Face is with just a few lines of Python code, you can load a model, pass in input text, and receive predictions for image generation and text classifications.
And to take it beyond that, if you want to succeed as an AI engineer and learn all the skills mentioned in this video in one place, the top resource I recommend is DataCamp. They offer comprehensive high-quality courses to help you build a strong foundation in AI engineering. The Associate AI Engineer for Developers track will actually teach you how to use OpenAI's API, Hugging Face, LangChain, and vector databases. You'll also master prompt engineering best practices, LLMOps principles for building and deploying AI systems, and creating more advanced AI applications. But if you come from a data science background and want to work more directly with foundation models, the Associate AI Engineer for Data Scientist track is definitely up your alley. This covers everything from machine learning fundamentals, deep learning with PyTorch, and using explainable AI techniques and responsible data management. Or if you're completely new to the field, the AI Fundamentals track is perfect for you. Here it offers a no-nonsense introduction to essential AI concepts. You'll learn core terminology and principles like how to use ChatGPT effectively, machine learning basics, and key ideas behind generative AI and its applications. What I love the most about DataCamp is the learning experience. Every course is fully interactive and hands-on. You're not just passively watching videos. For each of their tracks, they enable you to build real-world projects, which is the best way to learn. So, if you're ready to get started on your journey in learning AI engineering, be sure to check out the links down below in the description.
And now that we're proficient enough to know software engineering with AI and cooking with basic recipes, we really need to level up into layer three. And here we got intelligent systems for this one. So, instead of cooking and just following recipes, now we're going to learn how to create our own dishes. So, basically, instead of taking other people's AI workflows, we're going to learn how to construct our own. And for that I have four main tools/technologies we're going to cover.
LangGraph is one of the most important tools at this level. It allows you to build structured multi-step workflows around LLMs. So instead of a single prompt call like in ChatGPT, you can create logic like retrieve documents, evaluate confidence, call a second model, and then return output. So instead of treating AI just as a chatbot where you talk back and forth, you have now created your own AI system workflow with a chef's kiss. If you're interested, LangChain Academy has an introduction to LangGraph course absolutely for free. And for a starter project, I recommend building a simple multi-step workflow such as a research assistant that retrieves articles, summarizes them, and critiques its own output. And there are plenty of free GitHub examples and tutorials to really help you out.
MCP, Model Context Protocol, is also important when your models need structured access to tools. Think of this like a rule book in soccer. In soccer, you can pass the ball, shoot, and defend, but you can't suddenly pick up the balls with your own hands and make up your own rules mid-game. The rule book defines what actions are allowed and how they must be performed. MCP does the same exact thing for AI models. It defines what tools the model is allowed to use, what inputs those tools expect, and what outputs they return. For example, if you're building an AI assistant that can check order information from a Shopify database and send notifications through Slack, MCP has you clearly define the Shopify lookup function with the order ID and structured order data. And then it defines the rules for a Slack function with the channel and message to send. This prevents AI models from sending random messages and random formats to random channels.
RAG, or retrieval augmented generation, is best understood as giving the AI model an open-book exam instead of a closed one. So normally when you ask a large language model a question, it answers from what it remembers from training. And specifically ChatGPT, Claude, they're trained on the whole internet. So if you ask it a general question, it typically knows the answer to it. But if you're working at a company and they have a specific company handbook or specific internal policies, that's where RAG comes into the picture. When someone asks a question, the system first searches through your private documents to find the most relevant sections. That's the retrieval part. Then those retrieved sections are inserted directly into the model's prompt so the model can see them while answering. That's the augmentation part because you're augmenting the model's knowledge with fresh data. Finally, the model produces an answer using that supplied information. And that exactly is the generation part.
And if you're wondering how this works on a deeper level, this all comes down to our fourth concept of how AI models work, vector databases, which are basically smart storage systems. So say we have the employee handbook example, and you want to train an AI model to be able to answer any question about it whenever you have it. First, you would break the document into smaller pieces. This process is called chunking. Then each piece is turned into something called an embedding, which is basically a way of converting text into numbers so a computer can compare meaning instead of exact words. Those embeddings get stored inside of a vector database. So when someone asks a question, the system does not need to reread the entire document. Instead, it searches the database for chunks that are most similar in meaning to the question. For example, if someone asks "how many vacation days are allowed according to the handbook?", the system looks for chunks related to ideas of like vacation or PTO or leave policy. It then retrieves the most relevant section and passes only those to the model. The model then generates an answer based on those retrieved sections. And when you scale this up, it becomes incredibly powerful. Instead of guessing across thousands of documents, the AI system, by what you put, can search intelligently, retrieve only what matters, and respond accurately using real data. And that is why RAG and vector databases are foundational for enterprise AI systems.
Now, we should be really confident in our skills. And for layer four, we have scale without breaking. And at this stage, you're not just designing dishes anymore. You are running the entire kitchen while customers are ordering non-stop. It is no longer about just making your AI workflow operate, but rather your workflow must operate very well every single time without breaking. And if you're trying to become a proficient AI engineer, there are three tools you need to scale your AI workflows.
First, we got Docker. So, I want you to think of Docker like packaging Oreo cookies. When Oreo makes their cookies, they don't just bake them and throw them loose onto trucks randomly. They package them in identical sealed containers so every store receives the same product. Docker does the same thing for your AI applications. It wraps up your code, your dependencies, your models, everything into one consistent container, so it runs the same on your laptop, on a co-worker's laptop, and even on the cloud. Without Docker, your system might work for you, but fail for someone else.
Then we got AWS and GCP. So think of these cloud deployments like opening up multiple restaurant locations. Your kitchen at home can serve, say, maybe like five, 10 people. But what if 5,000 people show up? Cloud platforms like AWS allow you to host your AI system so anyone in the world can access it. This is where your chatbot, RAG system, or AI assistant becomes a real product instead of just a local experiment on your computer.
Then comes Redis caching, and this is all about efficiency. So think of it like keeping a frequently used ingredient on the counter. Like as a chef, it would be illogical for you to keep salt inside the pantry. Instead, you're going to keep it close to you because you're going to use it many times throughout the day and you don't want to waste time going back and forth to the pantry. Redis works the same way for AI systems. If thousands of users ask the same or very similar questions, you do not need to call the large language model every single time and pay for it repeatedly. You store the response once, keep it easily accessible, and reuse it when the same request comes in again. And trust me, this will dramatically lower costs.
The final layer to becoming a really good AI engineer is layer five, which is strategic AI operations. At this stage, you're not in the kitchen anymore. You're not cooking dishes, but rather you're looking at the entire restaurant from above. Are customers happy? Is overall food quality good? Are we making money? That is the mindset shift of layer five. In AI engineering, this is LLMOps. It is not just about whether your model works. It is about whether it delivers value to users sustainably. And for this we will cover three main areas.
First is evaluation frameworks like DeepEval. These tools are like food critics who consistently rate your dishes. You test your AI workflow for hallucinations and consistency. And to put this into practice, try integrating DeepEval into a small RAG pipeline and compare how different chunking strategies affect answer accuracy.
Next is analytics tools like PostHog or Amplitude. So imagine owning a restaurant and not knowing why some customers never return. That would be very dangerous and you could be losing a lot of money in business. So you need some way to track user experience, and analytics tools tell you exactly how users interact with your AI system. Which features do they use the most? Where do they drop off? Then you can use this data to improve your AI workflows.
Finally, one of the most important parts of being not just a good but great AI engineer is understanding cost governance and model routing. At scale, costs to run your AI models can explode. You need systems to decide when to use a large, expensive model versus a smaller, cheaper one. For example, for simple coding tasks, you might be able to use something fast like Claude Sonnet. But then for heavy reasoning tasks, you might need to use something like Claude Opus. This is like deciding which ingredients justify premium pricing and which do not. And learning this will save you a lot of time and money in the future and will holistically allow you to grow as the Michelin star restaurant owner of AI engineering to flourish.
Well, that's about all I have in this video. I really hope that you guys enjoyed it and if you did, make sure to hit the like button, subscribe if you haven't already. If you're interested in my absolutely free tech newsletter, the link for that will be down below in the description. And if you're interested in watching the interview that I had with the CEO of Microsoft, Satya Nadella, you might want to click this video right here.