Most roadmaps are written for someone with unlimited time and no other obligations. If you have five or ten hours a week around a degree or a job, following one of those plans produces the same outcome as following none: a long list of half-finished courses and no evidence of anything.
Courses are not the unit of progress
A course can be completed without learning, which is why completing one feels like progress and often is not. The unit that works is a check you can fail: something built, something measured, something that either works or does not.
Every stage below has one. If you cannot pass it, you have not finished the stage, regardless of how many hours or videos went in.
First 30 days: make something call a model
The goal is not understanding, it is contact. Write code that calls an LLM API, handles the response, and does something with it. Comfortable Python is enough — you do not need mathematics yet, and starting with linear algebra is the commonest way people stall before they begin.
- The check: a script you wrote that takes input, calls a model, and produces structured output you can use programmatically.
- Add: what happens when the model returns something that does not parse. That single question teaches more than the happy path.
- Skip for now: fine-tuning, vector databases, frameworks. All three are easier with context you do not have yet.
Days 31–90: one real project, finished
Pick something with a real data source and an actual failure mode. Retrieval over documents you care about is the usual choice because it is genuinely useful and it breaks in instructive ways.
The instinct will be to start a second project when the first gets difficult. The difficulty is the part with the value in it — retrieval returning the wrong thing, the model answering from the wrong section, cost climbing as the corpus grows. Those are the experiences interviews are about.
- The check: twenty questions with expected answers, and a number for how many your system gets right.
- Then: change one thing and see whether that number moves. That is evaluation, and it is the skill most candidates lack.
- Write the README as you go. Reconstructing your reasoning three months later is much harder than it sounds.
Months 4–6: depth, and being visible
By now you can build. The remaining gap is usually in two places: the operational side — cost, evaluation, what happens when it is wrong — and the fact that nobody knows you can do any of this.
Pick one of agents, retrieval quality or deployment and go deep enough to have opinions. Then make the work visible: the repository, the README, one written explanation of a problem you solved. Applying with evidence is a different experience from applying without it.
- The check: you can explain a decision you made and what it cost you, out loud, in two minutes.
- A structured programme with reviewed submissions helps here if self-direction is the part you struggle with.
What you can safely skip at the start
A great deal of the standard curriculum is genuinely important and genuinely not urgent, and confusing the two is why people stall in month one. You do not need to understand backpropagation to build a useful retrieval system. You do not need to have implemented a transformer to be employable in applied AI, which is where most current hiring is.
The material worth deferring until you have built something: the mathematics of optimisation, model architecture internals, training from scratch, and most of classical machine learning if your target is generative AI work. None of it is wasted — it is simply much easier to absorb once you have a concrete system to attach it to.
What is not optional, and is often skipped, is being comfortable in Python, being able to read an API reference, and being able to debug something that fails intermittently. Those three determine how fast everything else goes.
- Defer: optimisation maths, architecture internals, training from scratch.
- Do not defer: Python fluency, reading documentation, debugging intermittent failures.
- Revisit the deferred material once you have a system to attach it to. It lands much faster.
The two ways this usually goes wrong
The first is tutorial drift: finishing one course and immediately starting another, because starting is comfortable and finishing is not. The symptom is a folder of small completed exercises and nothing you would show anybody. The fix is a rule — no new material until the current check is passed.
The second is scope inflation: choosing a project ambitious enough to be impressive and then abandoning it at 40%. An abandoned repository is worse than none, because reviewers do open them. Halving your ambition at the start is almost always the right adjustment, and it is the one nobody makes voluntarily.
What this will not do
Six months of five hours a week is roughly 130 hours. That is enough to become genuinely useful and to have something real to show. It is not enough to become senior, and anybody promising a job at the end of a fixed number of weeks is describing a marketing claim rather than a plan.
How long it takes depends on where you start, how much of that time is focused, and the market you are applying into. The one thing consistently true is that people who finish one thing get further than people who start five.
Key takeaways
- 1.A course can be completed without learning. A check you can fail cannot.
- 2.First 30 days is contact, not understanding: write something that calls a model and handles a bad response.
- 3.The difficulty in your first real project is where the value is. Finishing it beats starting another.
- 4.Nobody can promise job-readiness by a date. Finishing one thing beats starting five.
Do it now, free
AI Career Roadmap Generator
Most roadmaps are a list of courses. This one is built around checks you can fail — a thing built, a thing measured — and it is scaled to the hours you actually have rather than the hours a roadmap wishes you had.
No sign-up. The internship is free to join too, with real project briefs and reviewed submissions.