rumman.ahmed

Architecture

The patterns behind the work

Four shapes I keep reaching for — because they keep holding. Each pattern links to the case studies where it earned its place.

Pattern 01

Event-driven systems

Most integration pain comes from systems calling each other directly: N sources × M consumers of coupling, failure that cascades, and every launch blocked on every dependency. Event-driven architecture inverts it — producers emit typed domain events onto a backbone, consumers subscribe and maintain their own read models. Nobody polls anybody.

The payoffs are isolation of failure (a slow upstream never takes your front-end down), independent deployability (new consumers subscribe without upstream change requests), and replayability — read models become disposable, rebuilt from history instead of migrated. I've shipped this shape at broadcaster scale (ABC's DigiStream pipeline) and museum scale (the Powerhouse Content API).

Event backbone — the canonical shape
Producers emit; consumers own their read models; the contract is the event schema.

Pattern 02

Headless, API-first platforms

When content or commerce lives inside the system that renders it, every new channel is a rebuild. Headless design separates the two: sources of truth stay specialised (CMS, collection system, CRM), a platform API exposes one stitched contract, and every 'head' — web, in-gallery, apps, partners — consumes the same thing.

The discipline that makes it work is treating the API contract as the product. Channels come and go; the contract compounds. At Powerhouse this is what lets a website, an exhibition interactive and a partner feed all render the same object without three integrations.

One contract, many heads
Specialised sources of truth behind one platform API; channels are replaceable.

Pattern 03

Multi-agent AI systems

A single AI chat session is a tool; a system of agents is infrastructure — and it needs the same architecture discipline as any distributed system. The shape I've converged on: an orchestration layer that routes work to specialist agents in parallel lanes, a queue that lets work proceed without a human present, and persistent memory so no session starts amnesiac.

The load-bearing decisions are the boundaries: agents own file domains so they never write over each other, autonomy is fenced by non-negotiable rails (no direct pushes to main, no secrets in commits, stop-and-explain on repeated failure), and every autonomous change lands behind the same CI and review gates a human's would.

Agent orchestration — myAI's shape
Deterministic rails around non-deterministic workers.

Pattern 04

Identity & auth architecture

Identity is where security posture and user experience meet, and both fail when every application solves login alone. The pattern: one authentication authority (an IdP like Auth0), every application reduced to a standards-based client, and — crucially — a clean split between authentication (who is this, verified) and the business system of record (who is this, to us).

Get that seam right and the rest follows: MFA and session policy set once, organisation-scoped access instead of per-app role tables, and new properties onboarding in hours because auth is configuration, not code.

One authority, many clients
Applications never see credentials; the CRM keeps owning the person.