The anatomy of an AI system

The course's whole architecture as one body. Like an anatomy book, each layer shows one system: switch them on together to see how dataflow, guardrails, risk, ownership, roles, cost, kinds of AI, and evaluation share the same skeleton. Every layer comes from a module; this page is the course read at a glance.

YOUR ESTATE AI PLATFORM · BOUGHT, RENTED, OR SELF-RUN PEOPLE Source systems records · documents Data platform lake · warehouse · pipelines Vector store chunk embeddings Language model reasoning · drafting Predictive models scores · forecasts Agent runtime plan · act · observe Staff and customers questions in Human approver actions signed off PERMISSIONS · BUDGETS · LOGS IN GATE OUT GATE ! ! ! ! ! ! ! ! OWNER OWNER OWNER OWNER OWNER OWNER SOURCE TEAMS DATA ENGINEER DATA ENGINEER AI ENGINEER DATA SCIENTIST SOFTWARE ENG SERVICE DESIGN ACCOUNTABLE ££ £ £££ £ ££ G G P G+R R ? ? ? ? Source systems: Where the organisation already keeps its records and documents. Nobody designed these to feed AI. Data platform: Lake, warehouse, and the pipelines between: where data settles and definitions are enforced. Vector store: Chunk embeddings, built to answer one question fast: what is nearest to this? Language model: Predicts the next token from everything in its window: fluent in general, weak on your specifics. Predictive models: Classical machine learning: learns from your history to score and forecast. Agent runtime: The plan, act, observe loop, choosing its path at runtime with tools. Staff and customers: Where questions begin: the people the whole body serves. Human approver: The valve on consequential actions: autonomy is earned through this box.
One body, eight systems. Combine layers to ask compound questions: risk plus ownership shows who owns which danger; cost plus kind shows where the money goes. Hover over, or tab onto, any component for its detail through the active layers; Escape dismisses.

Dataflow · the circulation

Pipelines feed the data platform from source systems; documents are chunked and embedded into the vector store, and history trains the predictive models. A person asks; the agent retrieves, the language model reasons, and anything consequential leaves only through the approver. Taught in Module 2 and Module 3.

Guardrails · the immune system

The platform runs inside permissions, budgets, and audit logs, with input and output gates on everything crossing the boundary and a human approval valve on consequential actions. Taught in Module 4.

Risk · the nervous system

Every component carries a signature risk: quality problems born in source systems, pipelines failing silently, stale or over-shared retrieval, hallucination and upgrade drift in the model, bias and decay in predictive models, actions beyond intent in the agent, shadow use among staff, and approval fatigue at the gate. Taught in Module 5.

Ownership · the skeleton

Each component bears a named owner senior enough to stop it; if naming one is difficult, that is the finding. The people boxes need no badge: they are the owners. Taught in Module 5.

Roles · who builds and runs it

Data and software engineers build the platform, the retrieval, and the agent. Data scientists build the predictive models, and that single box is the only place where the deep-learning frameworks named in job adverts, PyTorch and TensorFlow among them, are the actual day job. Everything that decides what to build, whether it works, and whether it may act belongs to people who write no code: the system owners, the service designers, the accountable owner, and the subject expert who can tell a good answer from a plausible one. Staffing the first group and forgetting the second is the commonest way an AI programme stalls after its demo. Each role is set out in full on who you need; the accountability behind them is taught in Module 5.

Cost · the metabolism

The language model burns most per call, the agent multiplies calls, checking rides on everything, and storage is cheap. The largest line sits outside the diagram entirely: the people who own, tune, and approve. Taught in Module 4 and Module 6.

Kind of AI · the tissue types

Generative tissue in the language model and retrieval; predictive tissue in the scoring models; rules in the pipelines and inside the agent's deterministic tools. Different tissues fail differently and are bought differently. Taught in Module 6.

Evaluation · the senses

The golden question set probes retrieval quality, model behaviour before and after upgrades, predictive outcomes across affected groups with drift monitoring, and agents end to end. What is not probed degrades politely. Taught in Module 5.

New here? The layers make most sense after the course: start at the overview, or look anything up in the glossary.