There’s a version of “AI engineer” that means you can call an API and parse the JSON. That version isn’t what will get you hired, and it isn’t what production systems need. The real job is different, more comprehensive and more end-to-end than you think.
It starts with data, because everything a model knows and everything it hallucinates traces back to what it saw and how that got tokenized. Then there’s the transformer itself: attention, embeddings, the training loop, the reason a context window has a limit and what happens as you approach it. On top of that is the application layer, where the model stops being a model and becomes a system with retrieval, memory, tools, retries, and state that has to survive a step failing halfway through. Then evaluation, governance, identity, and access, which nobody enjoys until the day they need it. And finally the layer where you explain the whole thing to a VP who controls the budget.
Most people learn these out of order. They start at the application layer because that’s where the fun is, get something working in a demo, and then call it a project!
So I created a complete FREE roadmap for AI Builders based on my experience building FDE team at Microsoft. I am sharing this roadmap in LIVE sessions from August 26 through September, all on Zoom, all recorded if you register and can’t make it.
For more details on how the roadmap works, read on!
Stage 1: Open the black box
Build an LLM From Scratch, September 3
Right now you call the model, and between your prompt going in and the tokens coming out there’s a region you have no real picture of, which makes every debugging session guesswork dressed up as intuition. So you build one: tokenizer, embeddings, attention, the training loop, small enough to hold in your head and real enough to run. You’ve cleared this stage when you can explain what your model is doing to a skeptical colleague, without hand-waving.
Stage 2: Learn the shape of agentic systems
Graph Engineering Explained Visually and Practically, September 10
If your experience is one-off chats and single-turn agents, you’ve seen about ten percent of what these systems do, and the other ninety percent is loops and graphs. The moment your agent needs to branch, retry, hand off, or hold state across steps, you’re doing graph engineering, whether or not you have a word for it. You’ve cleared this stage when you can sketch your agent on a whiteboard and everyone in the room follows it.
Stage 3: Give it something to know
Architect the Enterprise Context Layer for AI Agents, September 11
Your agent knows how to reason and knows nothing about how your business runs, so it guesses, and a confident guess about your refund policy is worse than no answer because someone will act on it. The fix isn’t more retrieval, it’s treating enterprise context as an architecture problem: the three walls teams hit scaling into production, and the components of a context layer that compounds in value. You’ve cleared this stage when you can name where your agent’s knowledge comes from and who owns keeping it correct.
Stage 4: Made Production Ready AI Agents
The 7-Factor Framework for Production-Ready AI Agents, September 11
The agent works and you’re still not sure you’d let it near the production database, which is the right hesitation, because a real agent moves money, updates records, and changes systems the business depends on. This one is with Workato: the 7-factor framework as an actual checklist for reliability, observability, and error handling, plus the orchestration practices behind production agent workflows. You’ve cleared this stage when you can look at any agent, yours or someone else’s, and give a defensible yes or no on shipping it.
Where this road continues. Stages 1 through 4 give you the mental models. If you want to build the thing rather than understand it, that’s the Agentic AI Bootcamp in Track A below.
Stage 5: Lock it down
Securing AI Agents: The 1Password Playbook, September 16
You probably can’t say how many AI agents are running inside your organization, and almost nobody can, including people whose job title suggests they should. They’re in browsers, in IDEs, on laptops, most authenticating with an API key someone pasted in months ago, and if that key leaks the attacker inherits every permission the agent has. We unpack it with 1Password, and you’ve cleared this stage when you have an answer to “which agent did that, and who let it?”
Stage 6: Get That FDE JOB
FREE Forward Deployed Engineer Interview Prep, September 15
FDE roles are among the hottest jobs in AI right now, with $1M+ packages at Anthropic, OpenAI, and Google, and most candidates prep for a standard SWE loop before walking into an interview that tests something else entirely. Strong engineers fail it with some regularity, not for lack of skill, but because they studied for a different exam. You’ve cleared this stage when you can walk into an ambiguous customer problem and scope it out loud, under pressure, without freezing.
Where this road continues. The free session tells you what the loop tests. Storytelling for Forward Deployed Engineers, in Track B below, is where you run the reps.
Where the road keeps going
The free sessions are the map. They’ll change how you think about agentic systems, and for a lot of people that’s the whole thing: take the models, apply them at work, move on.
If you’d rather do this live, with me and my team and a room full of people working on the same problems, there are two ways in.
The Agentic AI Bootcamp and certification, starting on September 16th, where you build the thing along with learning from my experience with AI transformation for the world’s biggest companies.
Use code EARLYBIRD for 30% off. Questions: contact@cloudgirl.dev
The Storytelling for FDE workshop on September 27, where I hand over what I’ve learned in 15+ years at Google and Microsoft, hiring Forward Deployed Engineers and sitting in rooms with customers
Use code EARLYBIRD200 for $200 off. Questions: contact@cloudgirl.dev
If you have questions and need to discuss with me send an email to me at this contact@cloudgirl.dev, also for my Substack community I am offering a 15 mins 1:1 call to answer any burning questions about these courses you can book it here.
