Enterprise Architecture
Complex enterprise AI solutions mapped to public architecture paradigms. Each covers the constraint, the architecture, what broke, and the measured outcome.
8 entries
A deterministic multi-agent platform deployed inside a client environment in under two weeks — sequential orchestration replacing open-ended prompt chaining, with full trace observability on every hop.
LLM-backed customer operations for a financial institution where data residency, auditability and deterministic validation are non-negotiable — and the model is the least interesting component.
A retrieval pipeline over 400k internal documents where permissions are enforced at retrieval time, not filtered afterwards — because post-filtering a leaked chunk is already too late.
A Terraform landing zone giving each tenant an isolated project with guardrails inherited from organisation policy — provisioned in minutes rather than the weeks a ticket-driven process took.
The unglamorous layer that decides whether an AI system survives contact with users — golden-set regression testing, trace-level observability, and cost attribution per conversation.
Replacing a rules engine that had accumulated 1,400 special cases with a hybrid pipeline — deterministic extraction where structure exists, agents only where judgement is genuinely required.
A production CI/CD reference implementation — Spring Boot to GKE via SonarQube and Trivy gates, Helm, and ArgoCD, with Workload Identity Federation removing every long-lived credential.
Scripted automation, then retrieval systems, then agents — each era moved the hard problem one layer down the stack. A reference architecture derived from what actually survived production.