Data & AI
Machine Learning Engineering
ML engineers sit between data science and software engineering. They take a modelling approach and turn it into a system that trains reproducibly, serves predictions under load, and keeps working as the data changes underneath it.
Start here
This path assumes you can already program properly and have some statistics. If you have one but not the other, fill that gap first — ML engineering is exactly the intersection, and being weak on either side shows immediately.
Before you begin
- Solid Python beyond scripting
- Understand basic statistics and evaluation
- Comfortable with Git and the command line
Engineering and ML foundations
Both halves. Being strong in only one is the commonest reason people stall on this path.
RequiredSoftware engineering for ML
Most ML code is written like a notebook and then has to be run like a service. That mismatch is the job.
Software engineering for ML
Most ML code is written like a notebook and then has to be run like a service. That mismatch is the job.
What to learn
- Modular Python and packaging
- Testing code that has randomness in it
- Dependency and environment management
- Git and code review
- Logging and configuration
Tools
- Python
- pytest
- Docker
- Git
Next: The modelling foundation.
RequiredMachine learning fundamentals
You will be asked to debug why a model is worse in production than in training. That is a fundamentals question, not a tooling one.
Machine learning fundamentals
You will be asked to debug why a model is worse in production than in training. That is a fundamentals question, not a tooling one.
What to learn
- Supervised learning
- Loss functions and optimisation
- Bias and variance
- Regularisation
- Evaluation and validation
- Data leakage
Tools
- scikit-learn
Next: Deep learning, where it earns its cost.
Deep learning
Powerful, expensive, and frequently unnecessary. Knowing when not to use it is part of the skill.
RequiredNeural networks
For text, images and audio there is no competitive alternative. For tabular data, gradient boosting usually still wins.
Neural networks
For text, images and audio there is no competitive alternative. For tabular data, gradient boosting usually still wins.
What to learn
- Layers, activations and architecture
- Backpropagation conceptually
- Training loops
- Overfitting and regularisation
- Transfer learning
- Reading a training curve
Tools
- PyTorch
- or TensorFlow
Practice
Next: Architectures for your domain.
RecommendedDomain architectures
You do not need all of these. You need the ones for the data you actually work with.
Domain architectures
You do not need all of these. You need the ones for the data you actually work with.
What to learn
- Convolutional networks for images
- Transformers for sequences
- Embeddings
- Fine-tuning a pretrained model
Tools
- Hugging Face Transformers
Project
intermediateA fine-tuned model
Take a pretrained model, fine-tune it on a dataset relevant to a real problem, and evaluate it honestly against a simple baseline. Report both.
- PyTorch
- Hugging Face
- MLflow or Weights & Biases
You know whether the deep learning approach actually beat the simple one, and by how much.
Next: Making training reproducible.
Reproducible training
A result you cannot reproduce is not a result.
RequiredExperiment tracking and versioning
Six weeks later you will need to know exactly which data and which parameters produced the model in production.
Experiment tracking and versioning
Six weeks later you will need to know exactly which data and which parameters produced the model in production.
What to learn
- Tracking runs, parameters and metrics
- Data versioning
- Model registries
- Seeding and determinism
- Comparing experiments honestly
Tools
- MLflow
- Weights & Biases
- DVC
Next: Features.
RequiredFeature engineering and pipelines
Training/serving skew — features computed differently in the two places — is one of the most common and most confusing production failures.
Feature engineering and pipelines
Training/serving skew — features computed differently in the two places — is one of the most common and most confusing production failures.
What to learn
- Feature pipelines as code
- The same transformation in training and serving
- Feature stores in outline
- Handling missing features at inference
Practice
Next: Serving.
Serving models
Where machine learning becomes a production system with users and latency budgets.
RequiredInference and APIs
A model behind a slow, fragile endpoint will not be used, regardless of its accuracy.
Inference and APIs
A model behind a slow, fragile endpoint will not be used, regardless of its accuracy.
What to learn
- Wrapping a model in an API
- Batch versus real-time inference
- Latency and throughput
- Batching requests
- Model formats and runtimes
- Quantisation and distillation in outline
Tools
- FastAPI
- ONNX Runtime
- TorchServe or a managed endpoint
Next: Scaling it.
AdvancedScaling inference
GPU inference is expensive. Efficiency here is often worth more than a marginal accuracy improvement.
Scaling inference
GPU inference is expensive. Efficiency here is often worth more than a marginal accuracy improvement.
What to learn
- GPU versus CPU serving
- Autoscaling
- Caching predictions
- Cost per thousand predictions
- Load testing a model endpoint
Project
advancedA served model with a latency budget
Deploy a model behind an API with a stated latency target, load test it, and optimise until it meets the target. Document cost per thousand predictions.
- FastAPI
- Docker
- A cloud platform
- A load testing tool
You know your model's latency at the 95th percentile and what it costs to run.
Next: Keeping it working.
Life in production
Models degrade quietly. This stage is about noticing.
RequiredMonitoring and retraining
Unlike ordinary software, a model can fail without erroring. It just gets worse, and nobody notices for months.
Monitoring and retraining
Unlike ordinary software, a model can fail without erroring. It just gets worse, and nobody notices for months.
What to learn
- Data drift and concept drift
- Monitoring prediction distributions
- Ground truth collection and delay
- Retraining triggers
- Shadow deployment and A/B testing models
- Rollback
Next: Doing it responsibly.
RequiredFairness and responsible deployment
Models make decisions about people. Bias in the training data becomes bias in those decisions, at scale and with an appearance of objectivity.
Fairness and responsible deployment
Models make decisions about people. Bias in the training data becomes bias in those decisions, at scale and with an appearance of objectivity.
What to learn
- Measuring performance across groups
- Interpretability tools
- Documenting intended use and limits
- Human review for consequential decisions
- Privacy in training data
Tools
- SHAP
- Model cards
Where this leads
You do not have to pick one now. These are the directions this path opens up once you are working.
Continue your journey
The paths closest to this one. Skills overlap more than the job titles suggest.
Data Science
Use statistics and modelling to answer questions data alone cannot.
ViewMLOps
Make machine learning repeatable, deployable and observable.
ViewAI Engineering
Build reliable systems on top of models you did not train.
ViewData Engineering
Build the pipelines that make everyone else's data work possible.
ViewYou do not have to do this alone
Our programs are free, taught live, and built around the same progression. Join one and work through it alongside other people.