Blog
Working notes on architecture, cloud infrastructure and AI systems — written as things ship.
A front desk answered visitors from six different tools on a shared tablet. Seven weeks later it runs on one screen that stores none of their data: every tile reads its source of record, the network decides who signs in, and the design fails closed. The architecture, the day it broke, and what is not proven yet.
A staff-picks feature for a museum visitor app went from pitch to a production prototype the same afternoon, then from CMS schema to TestFlight and the Play internal track in about thirty hours. Prototype the contract, not the screens: one shape, four teams building in parallel, a source label on every response, and a shared release pipeline with no special path.
Assessing a multi-channel visitor platform against four real failure modes instead of a tooling wishlist. Three telemetry dialects, correlation that breaks at every boundary, alerts that fire into a portal, why the join with SaaS vendors has to run on business keys, and the order that matters: an alert that reaches a person, then correlation, then the panel.
A front-of-house dashboard built for a wall-mounted tablet and a flaky venue network: a backend-for-frontend that keeps every credential off the device, the ops team's own tools used without being called, and offline data that knows what to forget. Then the rules underneath every repository: decisions written before code, one path to production, infrastructure enforced by policy, and why a rule that is not checked is only a preference.
A high-traffic visitor platform accumulated five cache layers that each kept their own opinion of time. How one monotonic version stamp made them agree — why freshness is a property of a field, not an endpoint — and how to arrive warm at a checkout you don't control.
A content layer that answers before it's asked — one document as the source of truth, read across every channel the museum owns.
A control-plane seam that turns “which cloud” into a runtime detail — six methods behind one interface, so the platform stops caring where it runs.
The architecture of myAI — a git-versioned memory layer, a specialist agent fleet, an autonomous off-hours runner, and an MCP gateway that ties them together. How it works, why it's shaped this way, and where it could go.
The architecture of Connect — an installable support hub where a bug report becomes an AI-triaged, human-approved, autonomously-built fix with a live public timeline. The triage→fix loop, the safety model, and how a year of plans reshaped the ambition.