A friend recently called and asked me what skills they need to become an AI engineer in 2026 and what would get them hired. As usual, I drafted a roadmap for them and thought it would be valuable to share it with you as well. Before we begin, we need to answer one crucial question: What is AI engineering, exactly?

There is considerable confusion online about this topic. You’ll see some people show you a data science roadmap and claim this is an AI engineering roadmap. They’ll direct you to learn mathematics, algorithms, how to train models, machine learning, and extensive theory. However, this is not what companies are searching for because they don’t need another data scientist. They need a real AI engineer.

Let’s keep it simple. An AI engineer is someone who builds systems that use AI to solve real business problems. This means you are not building ChatGPT or training a data model. You are building an AI system. What’s inside an AI system? You’ll find the following components. You will be connecting AI models, such as OpenAI models from Hugging Face. You will be connecting the company’s data, including databases, files, and documents. You will connect the company’s tools and apps, including email and internal services, as well as interfaces where users interact with the AI system.

As you can see, you are connecting components in one place. Your primary responsibility is to ensure everything is correct, secure, fast, scalable, and cost-efficient. This is exactly your role as an AI engineer. If you are interested, I’ll show you exactly how to become one, step by step.

The first phase covers the foundations, where you will learn core coding and how to collaborate with others on AI projects. The second phase addresses LLM interaction, where you’ll learn how to communicate and integrate AI models into your apps. The third phase is substantial, covering how to build and run a complete AI system. The final phase focuses on building your portfolio, showcasing your work online, and securing a job.

Let’s dive into phase one, where we’ll learn the foundations. We will start with Python. Python is the primary programming language you must learn as an AI engineer. Almost everything you see in AI systems, including tools, libraries, and models, is written in Python. In real-world projects, we utilise Python to clean data, integrate APIs, deploy models, and automate pipelines. If you want to build any AI system, Python will be your foundation.

I would like you to focus on four key points. First, data structures and loops. Second, functions and modules for code reusability. Third, how to handle files, especially those in JSON format. Fourth, how to make API requests to connect your apps to AI models. As an AI engineer, everything is clear. You have to master one programming language, and that is Python.

Once you start coding with Python, you will need to save your files. If you store them on your PC, this is a risk because you might lose them. Instead, we’re going to store our Python files in GitHub to manage different versions, enable sharing when applying for jobs or collaborating with colleagues on projects, and allow you to work on your code anywhere you want.

I want you to learn how to create a repository, push your code, make commits with clear messaging, pull and merge changes, and create branches for new features. Learn all of this from day one to become a professional. That’s all for this phase. You learn how to code and how to share your code.

In the next phase, we will begin learning about AI. The first topic is large language models, also known as LLMs. We need to learn how to communicate with AI models. The first and most underrated skill is prompt engineering. Most people think prompt engineering is simply typing something into ChatGPT and hoping for a good answer. However, there is actually a skill behind it. It’s all about how you communicate with a model so it understands exactly what you want and delivers a tailored answer precisely as you expect.

This means you will provide detailed and clear instructions and context. You’re going to tell the model who it is and what its role is, and show examples of the results you want. Each time you get an answer from the AI, you’re going to review it and improve your prompts step by step. For example, I have a prompt that I’ve been refining over the past year, which I use for brainstorming, allowing me to get exactly what I want from any given model.

To be honest, each time someone tells me the results from ChatGPT are bad and there are many hallucinations, I ask this person to show me their prompt. Usually, it’s only one very short sentence. Garbage prompts will give you garbage results. As an AI engineer, you must learn prompt engineering to configure AI models precisely according to your business requirements.

If you want help learning all these skills, you need to start building your own projects. AWS Educate has free courses for AI engineering. For developers who want to integrate AI into real-world applications, I recommend the GEN AI Developer Certification track.

For individuals with a data science background, I recommend the Associate AI Engineer for Data Scientists track with DataCamp. It will help your transition into full AI engineering. You will cover machine learning fundamentals, deep learning with PyTorch, working with the Hugging Face large model library, and you will learn how to fine-tune models. You’ll learn MLOps principles and many other AI and data topics.

If you are a complete beginner, I recommend the AI Fundamentals track IN Datacamp. It provides a no-code introduction to core AI ideas. You’re going to learn basic AI terminology, how to use ChatGPT effectively, and what machine learning and LLMs actually do. This course will give you the fundamentals you need to start building real projects, which is the most important part of learning.

Next, I want you to learn about the OpenAI API. With ChatGPT prompting, you use your browser to interact with the AI. It’s good and fun, but the real magic starts when you use the model behind ChatGPT by using the OpenAI API. The API enables you to interact with the same model, with one significant difference: within your own app, website, or product. With this, you can build chatbots and assistants directly within your tools, making it very simple to do so. It’s just a few lines of Python. You send a prompt, you get a response, and you display it wherever you want.

To learn this, I want you to focus on setting up your API keys, sending structured prompts with rules and context (here again, prompt engineering matters), and being careful with handling costs by managing tokens. This skill will take you from just playing with AI to building something using AI.

Moving on, you’re going to learn how to use models from Hugging Face. As you build your AI systems, you’re going to notice that you cannot rely completely on OpenAI because they are closed source. This means you have no control over the model. You cannot see exactly how it works. You’re going to pay significantly for using tokens every time. If you are not using Azure, your company data will leave your environment, which is a significant issue for many businesses.

This is where Hugging Face comes into the picture. It is the biggest community library for AI models. There are already more than two million models available. The optimal part is that most of them are free. This means you can find a model for almost any problem you might encounter. This is exactly why I keep repeating myself. As an AI Engineer in the industry, we no longer have to train anything from scratch. All you have to do is find the right model for your business case and fine-tune it.

The significant advantage of Hugging Face is that you can download the models locally to your machine and deploy them anywhere you want. With that, you stay in control. You can use sensitive data with the model because nothing leaves your environment, and you also reduce costs. What I want you to learn is to explore models in the Hugging Face hub, use transformers in Python, try pipelines, and learn how to deploy a model where the data lives.

That’s all for this phase. You have learned how to communicate with AI and how to integrate both OpenAI models and open-source models. Now we’re moving to the advanced phase in our roadmap, where you’ll learn how to build a real AI system. We’re going to start by learning the LangChain platform.

Let’s say you can talk to AI models, you can write effective prompts, and you can build impressive demos. However, this is not enough to build an AI system because you need to connect everything together. This is where LangChain comes in. You can use LangChain to orchestrate the entire process, connecting all the necessary models, tools, memory, and building your business logic so that the AI can take multiple steps to complete a full task.

This is exactly what an AI system does. It’s not simply one prompt and one answer. You have to learn LangChain, and I want you to approach it with four steps. Understand the core building blocks like chains, tools, memory, and prompts. Then, connect different models using LangChain and attempt to build a mini workflow using LangChain to solve real-world problems. Learning LangChain is the first step in building AI systems, and it will transition you from someone who can perform demos with AI to someone who can build an AI system. This is where real AI engineering starts.

Moving on to the next skill, this is going to be an interesting challenge. We’re going to learn about RAG, also known as retrieval augmented generation. The issue is that all AI models are pre-trained using public data. The company’s data is protected and not available publicly, which means AI models have no idea about your company’s data. We have to somehow connect the company’s data to the AI model. Here comes the concept of RAG.

Here’s how it works. First, you take all the company’s data, PDFs, and files and store them in a vector database. All you have to do is turn text into something called embeddings. Embeddings represent your text with numbers. Then you load and store all those embeddings inside the vector database. Once a user asks a question, it turns into an embedding, and the system starts comparing and searching for the closest match using semantic search. Once it finds the relevant information, the LLM model generates a response.

It’s all about adding memory to the LLM model to utilise your real data instead of relying on what it was trained on. If you want to learn RAG, then focus on four skills. Understand the RAG architecture and workflow. Understand how semantic search finds relevant context. Learn about vector databases and embeddings. After the theory, build a RAG system using LangChain. You’re going to prepare a file, split the text, store it in a vector database, and then create prompts and chains. This topic is substantial, but it will make you ready as an AI engineer.

Moving on to the next skill, we’re going to learn about AI agents. We all use AI chatbots like ChatGPT. You write a prompt, and it returns text. This is not enough. Companies want more than a nice answer on the screen. They want an AI that actually accomplishes tasks. This is exactly why we have AI agents.

An agent first thinks, and then it takes real action. For example, it might talk to a database, update records, call an API, or trigger a workflow. It is excellent for automating many tedious tasks we normally do at work, such as reading incoming emails and responding to them, creating meeting summaries, and generating those tedious Jira and ServiceNow tickets. It is far more than just chatting with AI. It is an AI that actually performs work.

To learn how to work with AI agents, you can do the following things. Learn how to call APIs and connect to databases using AI agents. Learn how to build an agentic RAG to search for documents and how to use multiple agents that can work together. AI agents are the real solution for automations at companies.

The next skill is the MCP, or Model Context Protocol. AI agents can only take real actions, such as checking emails, querying a database, or calling an API, when they are connected to external sources. Here, there are two significant problems. First, if you connect your AI agents directly to the production database, this is very risky. The second issue is that we have many external systems, and you’ll end up writing and building connectors for each tool. This requires considerable time and effort to create new connectors each time a new system is connected to AI agents. If you want to learn more about MCP, see this article.

This is exactly why we have MCP. It fixes those issues. We add a safe and standard layer between your AI agents and your sources. This has many benefits. First of all, you can plug and play any system into your AI agents without creating new connectors each time. Using this layer gives you full control over how the AI interacts with your sources, allowing you to add policies to protect your external sources. Utilising MCP servers and protocols enables everything to run faster. You can connect many things while feeling safe connecting AI agents to your sources.

To learn MCP, I want you to understand how the host, AI agents, and MCP server communicate through these protocols. Learn how to connect databases and APIs using MCP, how to set permissions and access scopes, and how to audit and log all actions. This is crucial for developing AI systems that are both safe and efficient. You have to understand this technique because you’re going to use it to build your AI system.

Moving on to the next skill, this is truly important. You have to learn LLM ops. Let’s say you have built an amazing AI system. Everything is working solidly, and you have a perfect RAG. Now it’s time to go live with your users. If you think you can celebrate and go on a long vacation after that, you are totally wrong. If you do that, everything will go down quickly. The model will be slow, the cost will be high, and your users will start complaining about the AI and the poor results.

This means there are post activities after you go live with your AI system. You have to understand the full cycle of LLMs from idea to production. Learn how to track prompts, how to test your RAG and optimise it, how to set up CI/CD, and how to do monitoring and alerts. There are several activities you must undertake after production to ensure your users trust your AI system. Otherwise, your project will fail.

Those are all the skills you need to become an AI engineer. Now we’re moving on to the final phase, where we’ll prepare you for getting hired. Nowadays, the market is challenging, and one of the optimal things you can do is build your portfolio. You can do that by building projects and publishing them on GitHub.

There are many ideas for AI projects, but here are my three recommendations. The first project is to talk to documents. Build a RAG assistant to answer questions from PDFs and files, and include citations in the answers to ensure trust in the information. The second project is talk to data. The user asks a question, and an AI agent converts it to SQL. It generates a query and then returns the results in a chart. It is very similar to the service that Databricks currently offers called AI Genie. It allows you to interact with data directly, but you prefer to build it from scratch.

Project three is an AI email automation agent. The AI agent reads emails and sends replies using specific, clear rules. Build one of these projects and make it your portfolio. Write clean README files, add an architecture diagram, and maybe even record a short demo about your project. You’re going to show companies not only that you know the skills but that you can build something.

If you want to take it a step further, you can become certified as an AI engineer. Here you have some excellent options. The AWS AI PRAC is a strong one. You get it directly from Microsoft. Or you can get the Databricks Generative AI Engineer certification. If you are targeting a company that already uses Databricks, this certification is valuable. You can get certification from DataCamp’s AI Fundamentals. It is a nice starter if you are completely new and want to get your first certification. If you are targeting more data science roles, you can get the AI Engineer for Data Scientists Associate certification. If you want to stand out from the crowd, combine projects with certifications. This can really push you forward in this challenging market.

This is the full roadmap to becoming an AI engineer. This is what I recommended to my friends. As you can see, it’s not about data science and training models from scratch. It’s all about building an AI system that every company is currently hiring people to create. If you enjoy this type of content and would like to support me, please subscribe, like, and comment. This really helps the algorithm. Connect with me on X, where I post daily, and follow my newsletter, where I write weekly about my experiences and the things I have learned at work and school. Thank you for reading, and I look forward to seeing you in the next article.

Leave a comment