AI Engineer Roadmap — From Beginner to Hireable
AI engineer roadmap from beginner to hireable: machine learning, LLM apps and shipping a model behind an API. Kaidoro adapts it to your background and deadline.
This is a starting point — make it yours
Use this goal to build your own roadmap — tailored to you and starting fresh.
Maths and tooling
Enough maths to understand what the libraries are doing.
Get fluent with NumPy and Pandas~12h
Vectorised operations, broadcasting, indexing, groupby, joins. Replace every loop you can.
Done when: you reach for vectorised operations before a for-loop by instinct.
Learn the linear algebra that matters~14h
Vectors, matrix multiplication, dot products, eigenvalues. Enough to know what a layer does.
Done when: you can explain why a neural network layer is a matrix multiply plus a nonlinearity.
Learn probability and statistics~14h
Distributions, expectation, variance, Bayes' theorem, hypothesis testing. The basis of every evaluation metric.
Done when: you can explain why 99% accuracy can be worthless.
Understand gradient descent by implementing it~8h
Fit a linear regression with gradient descent in plain NumPy. No libraries.
Done when: you can explain what the learning rate does when it's too large.
Classical machine learning
Still the right tool for most tabular problems.
Build models with scikit-learn~12h
Regression, classification, the fit/predict pattern, pipelines.
Done when: you use
Pipelineso preprocessing can't leak into your test set.Learn to evaluate honestly~10h
Train/validation/test, cross-validation, precision, recall, ROC-AUC, and the bias-variance trade-off.
Done when: you can explain why accuracy is the wrong metric for imbalanced data.
Master gradient boosting~10h
XGBoost or LightGBM — what usually wins on tabular data. Feature importance and early stopping.
Done when: you beat a linear baseline on a real dataset and know why.
Deep learning
Build one from scratch before using a framework.
Write a neural network with only NumPy~16h
Forward pass, backpropagation, a training loop. The single most valuable exercise here.
Done when: it learns XOR and you can explain every gradient.
Learn PyTorch~12h
Tensors, autograd,
nn.Module, optimisers, the standard training loop.Done when: you can port your NumPy network to PyTorch and get the same result.
Train a CNN on real images~14h
CIFAR-10 or similar. Convolutions, pooling, data augmentation, overfitting and how to spot it.
Done when: your validation curve tells you when to stop.
Understand transformers and attention~16h
Self-attention, queries/keys/values, positional encoding. Build a tiny character-level model.
Done when: you can explain what attention computes without hand-waving.
LLM applications
Most AI engineering jobs are now this.
Build an app on an LLM API~12h
Prompting, structured output, streaming, handling failures and rate limits.
Done when: it degrades gracefully when the provider is down.
Build a RAG pipeline~16h
Chunking, embeddings, a vector store, retrieval, and grounding answers in retrieved context.
Done when: it answers from your documents and says so when it can't.
Evaluate an LLM system~10h
Build an evaluation set. Measure whether a prompt change actually helped rather than trusting a vibe.
Done when: you can prove one prompt beats another with numbers.
Ship a model
A model in a notebook is not a product.
Serve a model behind an API~10h
FastAPI endpoint, input validation, batching, sensible latency.
Done when: it returns predictions over HTTP with proper error handling.
Containerise and deploy it~10h
Docker, then a cloud host. Handle model weights and cold starts.
Done when: someone else can call your model over the internet.
Track experiments~8h
MLflow or Weights & Biases — parameters, metrics and artefacts per run.
Done when: you can say exactly which run produced your best model.