DDD + GenericML + Knowledge Graphs
A decision‑centric reference architecture for engineering and regulated domains: start from domain semantics, define decision problems, build reusable data vectors, keep humans accountable, and preserve provenance in a graph.
Core principles
Decision‑first design
Every artefact traces to a decision or decision‑relevant workflow.
DDD semantics are authoritative
Meaning lives in bounded contexts & aggregates — not in DB columns or ML features.
Adaptability by extension
Evolve by adding contexts/aggregates/schema versions — not rewriting everything.
Human authority & accountability
Models provide evidence, not unilateral control (unless explicitly authorised).
Provenance is first‑class
Trace: data → vector → model → ensemble → evidence → decision → outcome.
Data vectors in practice
Inputs are floats or float[] (time‑windows / distributions). Strings become one‑hot. Bools become 0/1 (or a range later).
The architecture (3 big blocks)
Domain Bounded Contexts
Project‑specific semantics (assets, telemetry, maintenance, safety, etc.)
↔
Decision Intelligence Context
Decisions, vector schemas, model packs, ensembles, evidence bundles
↔
Knowledge Graph Spine (Neo4j)
Provenance, multi‑hop queries, Graph‑RAG, schema evolution
Key move: keep domain modelling separate from decision intelligence while maintaining controlled integration.
Decision Intelligence “core aggregates”
This is the reusable part that shows up in every project.
- DecisionCase — a decision episode: context snapshot, recommendations, approvals, outcome links.
- VectorSchema — versioned definition of a reusable vector (features, units, transforms, leakage rules).
- VectorInstance — a generated vector for a specific decision context (usually arrays stored outside the graph; graph stores refs + hashes).
- ModelPack — deployable model + training snapshot + validation evidence + validity envelope.
- EnsemblePolicy — how multiple model packs combine (weighted, stacked, rule+model, regime‑gated, etc.).
- EvidenceBundle — what was shown to humans and why.
- OutcomeRecord — what happened next (for learning + governance).
Download the full reference architecture
The complete doc includes templates, governance rules, and sample Neo4j query patterns.
Download reference architecture (.docx)