Field guide · 6 patterns
Architecture patterns that survive contact with production.
Six reusable blueprints distilled from a portfolio of working systems — spanning regulated SaaS, AI-augmented analytics, and edge-to-cloud IoT. Each strips away the domain and keeps the design reasoning, so you can borrow the structure for your own build.
How to read this collection
Designs, not products
These blueprints are deliberately generic. Every client name, brand, dataset and piece of proprietary logic has been removed. What remains is the part worth publishing: the shape of each system and the reasoning behind it.
- 6
- Reusable patterns
- 30
- Worked use cases
- 5
- Systems analysed
- 100%
- Domain-anonymised
The collection
Six blueprints
The Modular Monolith with a Pure Domain Core
One deployable, rigid internal layering, and business rules isolated as pure, testable functions — simple to run today, ready to split into services tomorrow.
Governed AI: Deterministic-First, Advisory Intelligence
Keep the model out of the critical path. A deterministic core decides and acts; the LLM narrates, suggests and drafts — always behind provenance, guardrails, and a human approval gate.
The Pluggable Adapter Seam — Stub to Real
Every external capability sits behind one interface with a built-in implementation and a real one. Demo end-to-end on stubs; swap to production with a single registration.
Multi-Tenant Isolation & Tamper-Evident Audit
Defense-in-depth tenancy bound top-down from the token, never the request body — plus a hash-chained audit ledger that commits in the same transaction as the change it records.
AI Over Enterprise Data: Grounded Analytics & the RAG Flywheel
A tiered cascade answers most questions before the LLM is called. When it is, the prompt is grounded in live schema, the output validated and self-corrected, and every success compounds.
Edge-to-Cloud IoT: Closed-Loop Telemetry & Control
Telemetry flows through a single message bridge to a deterministic engine that closes the loop — decided, commanded, observed, verified — with a hardware safety override independent of the network.
Threads that run through all six
The recurring instincts
Determinism at the core
The riskiest logic — workflows, scoring, control decisions — is isolated as pure functions with no framework or I/O, so it can be exhaustively tested and audited.
Seams over commitments
External engines, AI providers, key stores and brokers all sit behind narrow interfaces. The default ships as a stub; the real thing is a swap, not a rewrite.
Trust derived, never declared
Identity, tenant and jurisdiction come from the verified session — never from client input — and every state change leaves a tamper-evident trail.