MLOps and Responsible AI

Translating a business goal into a learnable target, choosing offline and online metrics, and knowing when not to use ML.

Immutable datasets, content hashing, lineage tracking and reproducible splits.

Assumes: Framing an ML Problem

Logging parameters, metrics and artefacts; comparing runs and avoiding lost results.

Assumes: Data and Model Versioning

Serialisation formats, ONNX, containerisation, and promotion through staging to production.

Assumes: Experiment Tracking

Batch, online, streaming and edge serving; shadow deploys, canaries and blue-green releases.

Assumes: Model Registry and Packaging

Data, concept and label drift; PSI, KL divergence and KS tests for detection, and alert design.

Assumes: Deployment Patterns

Scheduled vs triggered retraining, online learning, and safe automated rollout.

Assumes: Monitoring and Drift Detection

Seeds, deterministic kernels, environment pinning, and the limits of exact reproducibility on GPUs.

Assumes: Data and Model Versioning

Unit tests for data and features, invariance and directional-expectation tests, and behavioural test suites.

Assumes: Reproducibility

Training and inference cost models, hardware selection, autoscaling and caching economics.

Assumes: Deployment Patterns

Demographic parity, equal opportunity, equalised odds and calibration — and their provable incompatibility.

Assumes: Classification Metrics

Sources of bias across the lifecycle, and pre-, in- and post-processing interventions.

Assumes: Fairness Metrics

Global vs local explanation, intrinsic vs post-hoc methods, model cards, and when explanations mislead.

Assumes: Model Interpretability

The epsilon-delta definition, the Laplace and Gaussian mechanisms, DP-SGD and the privacy/utility trade-off.

Assumes: Probability Inequalities

Training without centralising data, FedAvg, non-IID client distributions and secure aggregation.

Assumes: Differential Privacy

Risk tiers, documentation duties, auditability, and the shape of current AI regulation.

Assumes: Explainability and Transparency