Python for Artificial Intelligence Training Course

Participants can inspect, clean and document datasets before model development.

• CSV, JSON, SQL, and API data loaded into pandas DataFrames
• Schema checks identify invalid types, duplicate records, missing values, and ranges
• Data profiling quantifies distributions, outliers, cardinality, and category frequency
• Missing-data mechanisms assessed as MCAR, MAR, or MNAR assumptions
• Train, validation, and test partitions prevent information leakage during development
• Data dictionaries record field definitions, ownership, lineage, and quality limitations

Participants can use numerical and visual analysis to identify useful modelling signals.

• NumPy arrays, broadcasting, indexing, and vectorisation replace slow Python loops
• Descriptive statistics reveal central tendency, spread, skewness, and unusual observations
• Matplotlib and seaborn charts expose trends, relationships, and class imbalance
• Correlation matrices distinguish linear association from causal claims and confounding
• Grouped aggregations compare customer, transaction, product, and geographic segments
• Distribution plots guide transformations, capping rules, and feature engineering decisions

Participants can create repeatable feature transformations using scikit-learn pipelines.

• Numeric scaling uses StandardScaler, MinMaxScaler, and RobustScaler selection criteria
• Categorical variables encoded through one-hot, ordinal, frequency, and target encoding
• Date fields produce calendar, elapsed-time, seasonality, and recency features
• Text fields transformed with tokenisation, n-grams, TF-IDF, and sparse matrices
• ColumnTransformer applies distinct preprocessing rules to defined feature groups
• Pipeline objects bind transformations and estimators without training-data leakage

Participants can translate a business question into a measurable machine-learning task.

• Prediction targets defined from decisions, outcomes, timing, and available evidence
• Classification, regression, ranking, and clustering matched to business problem types
• Baseline models establish minimum performance before complex model development begins
• Cost matrices express false-positive and false-negative consequences in operational terms
• Success criteria combine technical metrics, adoption conditions, and process constraints
• Target leakage identified through future information, proxy fields, and post-event variables

Participants can develop and evaluate regression models for continuous business outcomes.

• Linear regression estimates continuous targets using interpretable coefficient-based relationships
• Regularisation with Ridge, Lasso, and Elastic Net controls coefficient complexity
• Mean absolute error, RMSE, and R-squared compared against decision requirements
• Residual plots detect nonlinearity, heteroscedasticity, outliers, and systematic bias
• Cross-validation estimates regression performance across alternative data partitions
• Prediction intervals communicate uncertainty around estimated values and decisions

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

• Logistic regression produces class probabilities and interpretable feature contributions
• Confusion matrices quantify true positives, false positives, and missed cases
• Precision, recall, F1-score, and specificity selected for stated business risks
• ROC-AUC and precision-recall curves compare classifiers under class imbalance
• Decision thresholds adjusted using workload capacity, costs, and intervention priorities
• Probability calibration checked with reliability curves and Brier score

Participants can train and compare tree-based ensemble models for structured data.

• Decision trees split observations using impurity reduction and stopping rules
• Random forests reduce variance through bagging and random feature selection
• Gradient boosting improves weak learners through sequential residual correction
• XGBoost parameters control learning rate, depth, subsampling, and regularisation
• Feature importance examined using permutation importance and SHAP values
• Partial dependence plots show model response across selected feature ranges

Participants can compare model candidates using controlled experiments and documented evidence.

• Stratified k-fold cross-validation preserves class proportions across validation folds
• Grid search and random search tune parameters within defined search spaces
• Validation results recorded with datasets, features, metrics, parameters, and timestamps
• MLflow tracks experiments, artifacts, model versions, and comparison results
• Statistical uncertainty considered before claiming one candidate model performs better
• Champion-challenger selection documents acceptance criteria and release decisions

Participants can identify meaningful groups and anomalies without labelled outcomes.

• K-means clustering partitions observations using distance minimisation and centroid updates
• Feature scaling and dimensionality reduction improve distance-based clustering quality
• Silhouette scores, inertia, and cluster profiles assess segmentation usefulness
• Hierarchical clustering reveals nested groups through dendrogram interpretation
• DBSCAN identifies dense clusters and noise without preset cluster counts
• Isolation Forest flags unusual transactions, records, and behavioural patterns

Participants can prepare time-series data and evaluate forecasts using time-aware validation.

• Time indexes, frequency checks, gaps, and duplicate timestamps prepared for forecasting
• Lag features, rolling statistics, and calendar variables capture temporal patterns
• Walk-forward validation preserves chronological order during model performance testing
• Seasonal naive forecasts establish interpretable baseline performance for comparison
• Forecast error measured with MAE, RMSE, MAPE, and weighted business costs
• Data revisions and forecast horizons documented for operational planning use

Participants can build, train and diagnose basic neural networks using PyTorch.

• Tensors, computational graphs, and automatic differentiation support neural network training
• Dense layers transform features through weights, biases, and activation functions
• ReLU, sigmoid, softmax, and loss functions selected by task type
• Backpropagation updates parameters through gradient descent and learning-rate schedules
• Batch size, epochs, regularisation, and dropout control training behaviour
• Training and validation curves diagnose underfitting, overfitting, and instability

Participants can develop image-classification workflows using convolutional neural networks.

• Image tensors encode height, width, channels, pixel values, and normalisation
• Convolutional layers learn local visual patterns through filters and feature maps
• Pooling, stride, padding, and receptive fields affect learned image representations
• Data augmentation applies rotation, cropping, flipping, and colour transformations
• Transfer learning adapts pretrained vision models to limited labelled image datasets
• Class activation maps inspect image regions influencing classification predictions

Participants can prepare text data and build classical natural language processing models.

• Text cleaning handles casing, punctuation, URLs, numbers, and encoding errors
• Tokenisation and stopword handling prepare documents for language model features
• Bag-of-words and TF-IDF represent documents as weighted sparse vectors
• Naive Bayes and linear classifiers classify documents, messages, and tickets
• Named entity recognition extracts people, organisations, locations, and dates
• Error analysis groups language failures by vocabulary, ambiguity, and label quality

Participants can use transformer models and embeddings for semantic language tasks.

• Transformer attention relates tokens through contextual relevance within input sequences
• Token embeddings represent semantic similarity for search, classification, and clustering
• Pretrained models selected by language, domain, latency, licensing, and task
• Sentence embeddings support semantic search using cosine similarity rankings
• Fine-tuning adapts transformer classifiers using labelled domain-specific text examples
• Evaluation sets test accuracy, latency, bias, and failure patterns after adaptation

Participants can design and evaluate generative-AI prompts for bounded business tasks.

• Large language models generate tokens from learned probability distributions and context
• System, user, and assistant messages define roles, instructions, and examples
• Few-shot prompting supplies representative input-output pairs for constrained tasks
• Structured outputs use JSON schemas, validation rules, and retry handling
• Prompt evaluation compares factuality, relevance, format compliance, and safety
• Prompt injection and data-exfiltration risks inform defensive application design

Participants can design a retrieval-augmented generation solution for grounded responses.

• Document chunking balances context coverage, retrieval precision, and source traceability
• Embedding indexes store semantic representations for nearest-neighbour document retrieval
• Vector search ranks relevant passages using similarity scores and metadata filters
• Retrieval pipelines assemble grounded context before language-model response generation
• Citation patterns link generated answers to source documents and passage identifiers
• RAG evaluation measures retrieval recall, answer faithfulness, and response usefulness

Participants can specify a deployable AI service with monitoring and release controls.

• FastAPI endpoints expose validated model inference through documented HTTP requests
• Docker containers package Python dependencies, model artifacts, and runtime settings
• Unit tests verify transformations, prediction interfaces, and expected error handling
• Continuous integration runs tests, linting, and package checks before release
• Model monitoring tracks latency, errors, drift, performance, and usage patterns
• Rollback plans define triggers, model versions, owners, and service restoration steps

Participants can assess AI risks and define proportionate governance controls.

• Bias assessments compare errors and outcomes across relevant protected groups
• Data minimisation limits personal data collection, retention, and access permissions
• Model cards document intended use, limitations, metrics, risks, and ownership
• Human review points address high-impact decisions, exceptions, and appeal routes
• Threat modelling identifies adversarial inputs, poisoning, extraction, and access threats
• EU AI Act risk categories inform control selection and documentation priorities

Participants can build, review and present a complete Python AI solution with documented data, model, evaluation, deployment and governance decisions.

• Problem statement defines users, decisions, constraints, benefits, and measurable success criteria
• Dataset assessment records quality findings, leakage risks, permissions, and preparation choices
• Python pipeline prepares features, trains selected models, and produces reproducible outputs
• Evaluation report compares baselines, metrics, errors, thresholds, and acceptance criteria
• Deployment design specifies API interfaces, monitoring indicators, ownership, and rollback actions
• Peer review challenges assumptions, risk controls, evidence quality, and implementation feasibility