Career roadmap
Machine Learning Engineer
Train, fine-tune, and serve models. Own the numbers: data quality, metrics, drift, and inference cost.
Before you start ML Engineer
- Python including NumPy
- School-level linear algebra and probability you can refresh
- SQL
- Git and the Linux command line
Maths & data handling
Read a model's maths without flinching, and clean data without leaking it.
Enough to read a loss function and know what a gradient does.
- Vectors, matrices, dot products
- Matrix multiplication shapes
- Derivatives and gradients
- Chain rule
The language every metric is written in.
- Distributions
- Expectation and variance
- Bayes' rule
- Sampling error and confidence
Most of the job is reshaping data before a model ever sees it.
- Joins, groupby, window operations
- Dtypes and memory
- Vectorising instead of looping
- Polars for larger-than-RAM work
The number one junior mistake: any statistic computed before the split leaks the test set into training.
- Fit transforms inside the pipeline, per fold
- Time-based splits for temporal data
- Target encoding done wrong
- Duplicate rows across splits
BuildA cleaned public dataset with a documented notebook: every transform justified, leakage checks shown.
Classical machine learning
Beat a baseline honestly. Most paid ML work is still tabular.
The baseline you must beat, and the model you must be able to explain.
- Coefficients as explanations
- Regularisation: L1 vs L2
- Multicollinearity
XGBoost and LightGBM still win most tabular problems. Learn them properly.
- Bagging vs boosting
- Key hyperparameters and what they trade
- Early stopping
- Categorical handling
One train/test split is an anecdote. K folds is evidence.
- Stratified K-fold
- Grouped and time-series splits
- Nested CV for tuning
- Reporting mean ± std
97% accuracy on a 3% positive rate is a broken metric, not a good model.
- Class weights vs resampling
- PR-AUC over ROC-AUC
- Threshold selection by business cost
Pick the metric from the decision the model feeds, not from a tutorial.
- ROC-AUC vs PR-AUC
- Precision/recall trade-off
- Calibration curves
- Regression: MAE vs RMSE vs MAPE
BuildA tabular model that beats a documented baseline, with cross-validated scores and a feature-importance write-up.
Deep learning
Train and fine-tune neural networks on a real GPU budget.
Tensors, autograd, and a training loop you wrote yourself at least once.
- Datasets and DataLoaders
- Optimisers and schedulers
- Checkpointing
- Debugging NaNs
Two families cover most of the field. Attention is the one to understand deeply.
- Convolutions and pooling
- Self-attention and multi-head attention
- Positional encoding
- Encoder vs decoder stacks
Full fine-tunes are rarely necessary. Adapters give most of the gain for a fraction of the memory.
Ch 84 — RAG vs Fine-Tuning- Transfer learning
- LoRA / QLoRA
- Dataset size vs overfitting
- When RAG is the better answer
Memory is the constraint that shapes every decision.
- Mixed precision
- Gradient accumulation
- Batch size vs learning rate
- Spot instances and checkpoint resume
BuildA fine-tuned open model on a domain dataset, with before/after metrics and the cost of the training run.
Production ML
A model nobody can serve is a hobby project.
Most models do not need an online endpoint. Pick the cheaper shape that meets the need.
- Batch scoring pipelines
- Online inference APIs
- Latency budgets
- ONNX export and quantisation
The classic production failure: features computed one way in training and another way at serve time.
- Shared feature code
- Feature stores
- Point-in-time correctness
- Shadow scoring to detect skew
A model you cannot rebuild is a liability.
- Experiment tracking
- Model registry stages
- Dataset versioning
- Pinned environments and seeds
Models decay silently. Monitoring inputs catches it before the business does.
- Input distribution drift
- Prediction drift
- Delayed labels
- Retraining triggers
Offline gains are a hypothesis until online traffic agrees.
- Shadow traffic
- A/B design and sample size
- Guardrail metrics
- Rollback plan
BuildA model served behind an API with latency, drift, and data-quality monitoring, plus a documented rollback.
Specialise
Pick one domain and go deeper than a generalist can.
NLP, vision, recommenders, or forecasting. One, not four.
- NLP and embeddings
- Computer vision
- Recommender systems
- Time series and forecasting
Needed once one GPU is no longer enough.
- Data vs model parallelism
- DDP and FSDP
- Communication overhead
One paper a week, one reproduction a quarter, beats a course a year.
- Reading order: abstract, results, method
- Reproducing a result
- Spotting an unreported baseline
BuildA reproduction of one recent paper, with your notes on what the paper left out.
ML Engineer tools on your CV
- PyTorch
- scikit-learn
- XGBoost / LightGBM
- pandas or Polars
- MLflow or Weights & Biases
- Docker
- A cloud GPU
What ML Engineer employers ask to see
- One model in production with monitoring, not just a notebook
- A fine-tune with honest before/after numbers and its training cost
- A paper reproduction repo
Deeper bar than AI engineering and the maths is not optional. Hired by product companies with their own data, fintech and e-commerce risk teams, ad-tech, and global capability centres. Most paid ML work is still tabular, not LLM.
Content last reviewed 2026-08-31. Guidance only — no institute or paid placement is endorsed anywhere in this book.