Oliver Ernster
Principal Python engineer and software architect: APIs, integration and platform modernisation; 28 years at the seams of systems and author of the Decision Architecture series.
I don't just write about architecture. I ship it.
The work, in depth
Two areas of work, four questions each: what was hard, what was chosen, what it costs and what proved it.
Applications
Shipped applications, led by PigeonPost (email, calendar, contacts) and ClearBudget (solvency forecasting), with Meridian reading feeds over MMSP, the specification written to carry them. Stellody plays a music library it is structurally incapable of writing to.
The hard problem
Webmail wants engagement, Outlook wants a subscription and the successors to RSS monetise attention. In budgeting, a monthly total can look safe while the month itself is structurally unsafe. In music, a player can damage the collection it exists to play: a commercially available one wrote tags back across 21 albums of a real library.
The decision
Local-first, structurally. PigeonPost keeps mail with your provider (no account, no relay, no telemetry); ClearBudget projects the month day by day; Meridian reads a pull-only protocol with calm written into the wire; Stellody puts the tag-writing surface out of reach of every module that can read a tag.
The trade-off accepted
The local cache is not encrypted at rest; the site says so and points to OS disk encryption. MMSP holds a 300-second poll floor, so real-time delivery is impossible on purpose. Stellody describes a damaged tag and never repairs one, so the repair is yours to make in a tagger you chose; it is Windows today.
The evidence
Installers for Windows, macOS and Linux behind a 100% coverage gate on the core logic. A conformance suite covers every normative statement in the spec. Zero writes to a music file is a structural test rather than a promise. Every one of them is a tool I use daily.
Decision Architecture
The four-book thesis and the two instruments that make it executable: Fulcrum and LatencyLab.
The hard problem
Organisational theory cannot be run: claims arrive in prose and reorgs get tested on real people.
The decision
Make it executable. Fulcrum scores structural health 0 to 100 with no randomness: same model, same number, every time.
The trade-off accepted
The units are abstract, not calibrated to your KPIs. And the model never says what to decide, only where decisions can live.
The evidence
Typical archetypes collapse 83.4 to 12.0 as they grow; well-designed ones hold 98.7 to 64.9. Falsifiable by design.
Everything else
The rest of the catalogue, reachable in two clicks.