The AI/ML roadmap I'd hand to my younger self
A staged path from Python and classic ML to transformers, evaluation and deployment, with the project at each stage that actually proves you learned it.
Start with the tooling nobody talks about
Before models, before papers: Python, NumPy, pandas, one plotting library, git, and the command line. Boring, and it is most of the job. Every "ML problem" I have shipped was 70% data wrangling and 30% modelling.
You are done with this stage when you can take a messy CSV, explain why half its rows are junk, and produce a chart that answers a real question.
Classic machine learning before deep learning
Linear and logistic regression, trees, gradient boosting, and the whole ritual around them: train/validation/test splits, cross validation, leakage, class imbalance, precision versus recall.
Nothing about this gets replaced by transformers. The evaluation instincts are the same, and they are what separate a working model from a demo. My stage project was a churn model with an honest cost analysis: how much money a false negative costs versus a false positive.
Deep learning, but small
PyTorch, tensors, autograd, a training loop written by hand at least once, embeddings, and overfitting on purpose so you can see what it looks like. Stay on small datasets. You are learning the mechanics, not chasing accuracy.
Then one vision task and one sequence task, both trained by you from scratch.
Transformers and the LLM layer
Attention, tokenisation, context windows, and how a modern decoder-only model is actually trained, at least at the level where the tradeoffs make sense. Then go up a level and use models through APIs: prompting, structured output, tool calling, streaming, cost and latency budgeting.
This is where most teams now live, and where the useful instincts are economic rather than mathematical: when to call a model, when to call code, when to call neither.
Retrieval, agents, and evals
Vector search, chunking, hybrid retrieval, reranking. Then agent patterns: tool loops, planners, memory, guardrails. And finally evals, which is the stage most people skip and the reason their systems silently regress.
Build one retrieval system and one agent, and write evals for both before you polish either.
Deployment
An API, a container, versioned data and models, and monitoring for drift. You do not need a platform team. You need one service that stays up and tells you when it stops being right.
How to tell you are progressing
Not by certificates. By scope: you stop asking "which model" and start asking "what would we measure, and what does being wrong cost us". That question stays the same for the rest of your career, whatever the architecture of the month is.
