Reference architecture for controllable industry intelligence

Design goals

Industry intelligence must operate under real network limits, imperfect data, permission boundaries and organizational processes. The system should be deployable, explainable, auditable, replaceable, extensible and operable over time.

Six-layer reference architecture

The layers are infrastructure and edge; data ingestion and governance; models and knowledge; spatial and business services; applications and collaboration; operations and security. Versioned interfaces prevent lock-in to one model or platform.

• Unified identity, access, secrets and audit
• Model gateway, RAG, vision services and rules
• GIS/asset model and business APIs
• Observability, evaluation, canary and rollback

Data and AI governance

Datasets, documents, prompts, models, rules and knowledge bases need owners, versions, scope, freshness and evaluation history. High-risk outputs require citations, confidence, human approval and tool allowlists.

Delivery and operations

Use environment manifests, locked dependencies, automated builds, separated configuration, backup/restore and runbooks. Acceptance covers function, performance, security, maintainability and training.

Evolution path

Begin with high-value, lower-risk, reviewable workflows, then expand data, models and organizational scope. Preserve baselines and exit mechanisms at every step.