The unit of analysis becomes the governed decision architecture.
An end-to-end AI strategy does not imply one central super-model. It requires a modular architecture in which models publish and consume data objects, explicit policies govern action, digital twins support scenario testing, and humans retain defined decision rights.
Separate data services, models, policies and execution systems.
Every AI output has a schema, timestamp, provenance and confidence.
Changes propagate through typed events rather than ad-hoc messages.
Risk determines whether AI recommends, requests approval or executes.
Log decisions, model versions, drift and realised outcomes.
High-impact changes are tested in a digital twin before execution.
LLM outputs retain retrieved evidence and source lineage.
Operational outcomes update models, rules and organisational routines.
A simple event-driven end-to-end pattern.
Decision service pseudocode
Teaching progression
| Stage | Student task |
|---|---|
| Problem | Define the operational decision and value-chain boundary |
| Representation | Select table, graph, vector, text, image or event schema |
| Model | Train/evaluate AI with uncertainty and baseline |
| Decision | Add optimisation, policy and human authority |
| Integration | Connect model output to a simulated workflow |
| Learning | Evaluate performance, drift and organisational consequences |
Research position
The team studies AI not only as an analytical technology but as an operational architecture: how data is represented, how models are programmed into workflows, how decisions cross organisational boundaries, and how industrial systems learn end-to-end.