The AI Engineer Map Drawn From 40,000 Job Ads: Code, Agents, RAG and Cloud

Baraa spent two months building a system that runs every day and collected over 40,000 data and AI job ads from more than 10,000 companies in 100 countries; more than 4,000 of those ads are hiring AI engineers. The role is brand new and platforms shift at a dizzying pace, so there is no single roadmap everyone agrees on and most roadmaps are opinions. The video’s mantra is clear: without data you are just another person with an opinion. That is why the entire narrative walks through real job ads and unpacks what companies expect from an AI engineer — skills, tools, salaries, day-to-day tasks — and adds the fast-rising forward deployed engineer role on top.

Who is an AI engineer? In the video’s framing it is a software engineer who builds applications with AI models, with a thin but real data-engineering layer that moves and feeds data into the system. It sounds simple until you open the job boards and meet the chaos. The biggest cluster is AI engineers, but the same job travels under many names: applied AI, gen AI, generative AI, agentic AI all map to the same core. Around it are smaller clusters: researcher, architect, product manager. And the forward deployed engineer, which did not exist two years ago and now curves upward. The naming noise shows the market has not yet normalized the concept.

The skill analysis method is disarmingly simple. Baraa scans the “what we are looking for” section and respects the internal ranking of each ad; most listings are long wish lists nobody can fully satisfy, the top lines are must-have and lower lines are nice-to-have. Frequency matters, but position matters as well. The video groups mentions into clusters so patterns emerge instead of counting tools one by one. The sponsored training segment is distilled to one neutral line here: an example program that starts with hands-on Python and then teaches LLMs, agents, RAG and operations together, with first chapters available to try for free.

Most-wanted skills and the core trio

The top of the list holds no surprise. Python is the undisputed number one, almost every job depends on it; second comes the ability to work with large language models. In the programming-languages cluster Python dominates and mentions of TypeScript and SQL stay marginal. The recommendation follows the data: master Python well into advanced topics and add SQL as a second language if you need one. The message is first engineer, then AI. Language choice is a tool decision, not the thesis. Python’s weight aligns with the thesis that an AI engineer uses models rather than trains them from scratch.

Inside the AI-skills cluster a core trio separates: LLMs at the top, followed by AI agents and RAG. The video gives crisp functional definitions: the LLM is the model inside your application, an agent connects that model to tools so it takes actions, and RAG searches company documents and feeds relevant context to the model so it stops hallucinating. What comes after this trio corrects a common misconception: you do not need to train models from scratch, nor a PhD or deep math. That is the data scientist’s job. The AI engineer’s job is to build the system; the scientist discovers the model. Once that boundary is clear, the path simplifies.

For tool and platform choices the video argues for simplification. On the model side one strong LLM is enough, GPT or Claude, Baraa uses Claude daily but switching is normal; PyTorch and TensorFlow can stay aside for now because they are about training, which is not the core job. In the framework and ecosystem corner LangChain and LangGraph stand out, with OpenAI APIs and Hugging Face plus vector databases around them. On clouds AWS leads, followed by Azure and Google; do not try to learn all three, pick one, the general nudge is AWS with a note that Europe leans more toward Azure. The decision should be adjusted to the local market.

The deployment skill cluster is the video’s most emphatic note. Building the app does not end the job, you must know how to ship it to the target environment. The list starts with APIs, then CI/CD and Docker. Baraa flags the most common mistake he sees: candidates learn how to build and skip how to deploy. The line between someone tinkering at home and an engineer is drawn exactly here. Moving from a local machine to a production environment is the differentiator. Deployment is not a nice-to-have, it is the career signal.

Day-to-day tasks, pay, and the new role’s anatomy

After skills comes the more honest part: what you will actually do every day. The same rank-aware reading is applied, this time to the “what you will be doing” section. Number one is building LLM applications, number two is building AI agents for process automation — not chatbot demos but systems that follow a plan, call tools and finish work. Number three is deploying to the cloud and watching it, if production breaks you fix it. Number four, and the hardest, is evaluation: measuring results with benchmark sets and scores. Calling a model is easy, answering whether we can trust the result is hard; that is where the edge forms.

Beyond the headline numbers there are the tasks that complete the day: cross-team collaboration, integrating the model with existing systems, maintaining backends and APIs, keeping RAG alive, and handling safety and compliance. Baraa summarizes the day in one line: build the AI and the software around it, deploy it, watch it, measure whether it can be trusted. Those four verbs are the skeleton of the AI engineer’s job description. The video presents that skeleton with equal weight on frequency and responsibility; deploy and watch are as legitimate as build.

Next come two sensitive topics: pay and credentials. Baraa warns up front that numbers come only from the United States and only about 30 percent of US ads mention salary, and the picture shifts dramatically by country. For a sense of scale he still shares US averages: around 150k for juniors, 200k for seniors, 235k for leads. Compared with data engineers, AI engineers sit higher at every level and companies are willing to pay a premium to become AI-driven. On degrees half the ads do not even mention one, so entry without a degree is realistic though a degree helps odds. Certificates are even fainter, mentioned in fewer than 3 percent of ads; chasing badges is not the game, an end-to-end portfolio project carries far more weight and gives you a story to tell in interviews.

Now to the promised close cousin: the forward deployed engineer. Short definition: solution engineer for AI. The industry problem is familiar, hundreds of projects are built quickly but die quickly because they never integrate with core systems, burning money and time. The FDE closes that gap: listening to customer requirements and walking the team step by step toward an integrated AI system, owning the whole journey. The bank example makes the difference clear: a bank hires an AI engineer to build inside, while platform vendors like Databricks, OpenAI or Anthropic send their FDE to the customer site so the platform actually works there. The trend accelerates on the vendor side.

FDEs and AI engineers share the same skills, the lifestyles diverge. The video’s checks are striking: more than 80 percent of FDE ads include a task about the customer’s own systems, half mention travel, and former solution engineers are frequently requested. Both build the same thing, but one builds inside the company and the other lives on the customer site with the platform. The choice is a career preference about where you want to build. An AI engineer builds directly inside, an FDE integrates outside.

The roadmap now condenses all of this evidence into a clear phased plan. Phase one: truly master Python, not just basics but advanced topics. Right after that learn APIs and Git plus CI/CD; an AI application is ultimately a service that other software will call and your system must respond. From that foundation move to the AI core: first understand LLMs, what they can and cannot do and how to control their behavior, then agents where the model calls tools, follows a plan and completes tasks. Next comes RAG: build a vector database, feed it with company data so the model answers from documents instead of making things up. And the ring that must not be skipped: evaluation, proving that the system you built can be trusted.

Phase three is tool and platform decisions: pick LangChain or LangGraph, think of it together with an AI provider and vector layer, for clouds the general drift is AWS but check local practice. The finale is the most important: an end-to-end portfolio project built around a real use case. Beyond demo chat or PDF question-answering: a RAG plus agent ensemble that answers from documents and takes actions, ideally served via FastAPI, measured with a 15-question evaluation set, compared by changing prompts and scoring again, hardened with refusal rules and speed and safety layers. The closing step is deployment: a container, a pipeline that tests every change, one cloud and a week of live log watching. The story you will tell in interviews will not be badges but this end-to-end care. The video’s closing note is balanced: AI engineering is among the highest-paid engineering tracks and does not require a PhD, but because tools age fast the biggest risk is that what you learn goes stale; the map must stay alive through continuous updates.