The architecture choices that recur.
Seven interactive trees for the decisions that come up most often in real engagements. Each walks the — specific — conditions that make a different answer right, ends in a recommendation with trade-offs and watch-outs, and links to the matching diagnostic, reference architecture, anti-pattern, or essay. The path is captured in the URL hash: bookmarkable, shareable.
Build or buy an AI gateway.
Use-case count × regulatory posture × platform-team funding × sovereignty constraints. Skip-vs-buy-vs-build sized to the conditions.
Open the tree →RAG, fine-tune, prompt — or hybrid.
Knowledge-vs-behaviour fork. Knowledge change cadence. Example count for behaviour. Strict-vs-loose for hybrid. Where the “just fine-tune it” reflex goes wrong.
Open the tree →OPA, Kyverno, Cedar — which engine, where.
Picked by enforcement surface (admission · CI · IaC · app-authz), not by language preference. Where the Rego learning curve flips the decision.
Open the tree →Shared, cell-based, or single-tenant.
Uniform load × isolation requirements × head-tenant count. The migration triggers (when to add cells, when to extract a single-tenant stack).
Open the tree →Kubernetes, managed PaaS, or FaaS.
Service count × traffic shape × platform-team funding × multi-cluster needs. Where K8s adoption without a platform team becomes a tax.
Open the tree →Sync, async, or event-driven.
Does the caller need the answer? Cross-org vs internal? Queue vs event-stream? The hybrid pattern (sync critical-path + event for the rest).
Open the tree →Monolith, modular monolith, microservices.
Team count predicts the right shape more than tech fit. Where the modular monolith beats both extremes. Strangler-fig extraction at scale.
Open the tree →