Artificial Intelligence and Machine Learning Applications Training Course

Participants can prioritise AI opportunities using value, feasibility, risk and readiness criteria.

• Use-case canvas defining users, decisions, actions and measurable benefits
• Problem statements converted into target variables and prediction horizons
• Value-effort matrix comparing candidate AI and automation initiatives
• Data readiness assessed through availability, quality, access and lineage
• Risk screening identifies fairness, privacy, security and safety concerns
• Prioritisation scorecards support portfolio decisions and stakeholder alignment

Participants can specify the data required for a machine learning solution and identify quality risks.

• Observation units, labels, features and identifiers defined precisely
• Training, validation and test datasets separated without information leakage
• Data lineage traced from source systems through model-ready tables
• Missingness mechanisms assessed before imputation or record removal decisions
• Outliers examined using business rules, distributions and domain evidence
• Data dictionaries document field definitions, provenance and permitted uses

Participants can use Python notebooks to load, inspect and transform analytical data.

• Jupyter notebooks organise code, outputs, assumptions and narrative findings
• Pandas dataframes filter, join, aggregate and reshape source records
• NumPy arrays support numerical transformations and matrix-based calculations
• Functions package repeatable data preparation and evaluation tasks
• Versioned environments record package dependencies and reproducible execution
• Readable code separates configuration, processing, modelling and reporting

Participants can investigate data distributions, relationships and quality issues before modelling.

• Univariate distributions reveal skew, spread, sparsity and implausible values
• Cross-tabulations compare outcome rates across operationally relevant segments
• Scatterplots expose nonlinear relationships, clusters and influential observations
• Correlation analysis identifies redundant predictors and possible multicollinearity
• Target leakage detected through timing, provenance and feature availability checks
• Data quality findings translated into remediation priorities and modelling constraints

Participants can construct model-ready features while preserving reproducibility and preventing leakage.

• Categorical variables encoded using one-hot, ordinal and target encoding
• Numerical features scaled, transformed and binned for model requirements
• Date fields converted into lags, seasonality and recency measures
• Text fields transformed using tokenisation, TF-IDF and embeddings
• Pipeline objects apply identical transformations during training and scoring
• Feature stores considered for governed reuse of production features

Participants can build and interpret regression models for continuous-value business outcomes.

• Linear regression estimates continuous targets with interpretable coefficient effects
• Regularisation using Ridge and Lasso controls unstable coefficient estimates
• Tree-based regressors capture nonlinear effects and feature interactions
• MAE, RMSE and MAPE compared against operational error tolerance
• Residual plots identify bias, heteroscedasticity and unmodelled structure
• Prediction intervals communicate uncertainty around individual model estimates

Participants can build classification models and set decision thresholds for operational use.

• Logistic regression estimates probabilities for binary decision outcomes
• Decision trees provide transparent rules for initial classification baselines
• Random forests and gradient boosting improve predictive classification performance
• Confusion matrices quantify false positives, false negatives and correct decisions
• Precision, recall and F1 score matched to intervention priorities
• Threshold selection reflects capacity, error cost and customer impact

Participants can evaluate model reliability, explain predictions and identify generalisation risks.

• K-fold cross-validation estimates performance variation across data partitions
• Stratified sampling preserves class prevalence during training and validation
• ROC-AUC and precision-recall curves assess ranking quality under imbalance
• Calibration curves test whether predicted probabilities match observed frequencies
• SHAP values explain feature contributions for individual and aggregate predictions
• Model cards document intended use, limitations, metrics and approvals

Participants can handle rare-event prediction problems using appropriate sampling and decision methods.

• Class imbalance diagnosed through prevalence, confusion matrices and baseline metrics
• Random undersampling and oversampling compared against information-loss risks
• SMOTE generates synthetic minority examples within feature-space neighbourhoods
• Class weights penalise costly minority-class errors during model training
• Precision at k supports constrained investigation and intervention capacity
• Cost matrices convert operational consequences into threshold selection criteria

Participants can identify meaningful groups and anomalies when labelled outcomes are unavailable.

• K-means clustering groups observations using distance-based similarity measures
• Hierarchical clustering reveals nested structures and alternative cluster counts
• Silhouette scores assess cluster separation and internal cohesion
• Principal component analysis reduces dimensionality for visual inspection
• Isolation Forest identifies anomalous records without labelled fraud cases
• Segment profiles connect statistical clusters to service and commercial actions

Participants can prepare time-series data and assess forecasts for planning decisions.

• Time-indexed data prepared with consistent frequency, gaps and cut-off dates
• Trend, seasonality and cyclicality separated through decomposition analysis
• Lag features capture autocorrelation and delayed operational effects
• Rolling-origin validation prevents future information entering historical forecasts
• ARIMA and gradient boosting compared for forecast accuracy
• Forecast bias monitored against inventory, staffing and budgeting consequences

Participants can design NLP workflows for classification, extraction and document-based analysis.

• Text preprocessing applies normalisation, tokenisation, stopwords and stemming choices
• TF-IDF vectors represent document terms for baseline classification models
• Named entity recognition extracts people, organisations, locations and dates
• Sentiment classification distinguishes opinion signals from operational categories
• Topic modelling identifies recurrent themes within large document collections
• Human review queues handle low-confidence classifications and sensitive content

Participants can assess when neural networks are appropriate and configure a basic deep learning model.

• Neural network layers transform inputs through weighted nonlinear activations
• Backpropagation updates weights by minimising a specified loss function
• Training, validation and test loss diagnose underfitting and overfitting
• Dropout and early stopping reduce overfitting during model training
• Convolutional networks support image classification and visual inspection tasks
• Compute cost, data volume and explainability shape architecture selection

Participants can specify generative AI applications, prompts and controls for business tasks.

• Transformer architectures generate text using attention across token sequences
• Prompt templates define role, context, constraints, examples and output format
• Few-shot prompting improves task reliability through representative demonstrations
• Temperature and top-p settings control response variation and determinism
• Hallucination risks assessed against source availability and decision consequences
• Content filtering addresses harmful outputs, sensitive data and policy violations

Participants can design a retrieval-augmented generation solution for controlled organisational knowledge access.

• Document chunking balances retrieval precision, context continuity and update frequency
• Embedding models convert text passages into semantic vector representations
• Vector databases retrieve relevant passages using similarity search
• Grounded prompts require answers to cite retrieved source material
• Retrieval quality evaluated using relevance, coverage and citation correctness
• Access controls restrict documents according to user permissions and classifications

Participants can plan deployment and monitoring arrangements for a production machine learning service.

• Batch, real-time and embedded scoring selected by decision timing
• REST APIs expose model predictions to applications and workflow systems
• CI/CD pipelines test code, data schemas and model artefacts
• Model registries control versions, approvals and promotion between environments
• Data drift and concept drift trigger investigation and retraining decisions
• Service-level objectives define latency, availability and prediction-quality expectations

Participants can assess AI risks and define proportionate governance controls for an AI system.

• NIST AI RMF functions organise governance, mapping, measurement and management
• Impact assessments identify affected groups, harms and accountable owners
• Fairness metrics compare selection rates, errors and outcomes across groups
• Privacy by design limits collection, retention, access and secondary use
• Threat modelling addresses prompt injection, poisoning and model extraction
• Independent validation challenges assumptions, controls, performance and documentation

Participants can define the operating model, costs and accountabilities required to sustain an AI solution.

• Product ownership assigns decisions, backlog priorities and benefit accountability
• RACI matrices clarify responsibilities across business, data, technology and risk
• Build-buy-partner analysis compares capability, cost, control and supplier dependency
• Total cost of ownership includes data, compute, licences and support
• Benefits realisation plans track adoption, process change and realised value
• Model retirement criteria address obsolescence, replacement and records retention

Participants can build, review and present a complete AI solution proposal with model, governance and implementation evidence.

• Select a decision problem with measurable value and accountable owner
• Prepare a model-ready dataset and document material quality limitations
• Build and compare baseline machine learning models using defined metrics
• Justify threshold, workflow integration and human-review design choices
• Complete model card, risk assessment and monitoring plan documentation
• Present implementation case, assumptions, controls, costs and expected benefits