diff --git a/content/work/api-developer-portal.mdx b/content/work/api-developer-portal.mdx
new file mode 100644
index 0000000..cc118bc
--- /dev/null
+++ b/content/work/api-developer-portal.mdx
@@ -0,0 +1,56 @@
+---
+title: "A developer portal for a bank's partner APIs"
+summary: "The approach for opening a bank's APIs to business partners: onboarding, sandbox, go-live approval and fees, and a review of the bank's own high-level design."
+sector: "Banking"
+role: "Solution architect"
+themes: ["architecture-integration", "domain-modeling-rules"]
+kind: "solution-design"
+period: "2026"
+date: "2026-08-10"
+stack: [".NET", "React", "Keycloak", "API gateway", "PostgreSQL"]
+---
+
+## Context
+
+A bank wanted to open its APIs to business partners through a developer
+portal: partners sign up, try the APIs in a sandbox, get certified, go live and
+are charged for use. Behind it sat an existing API gateway and the core banking
+system.
+
+I worked through the bank's requirements, about 165 of them, mapped each to a
+part of the solution, proposed the approach and the target architecture, and
+reviewed the bank's own high-level design against them.
+
+
+
+## What I proposed
+
+- **One approval component.** Partner sign-up, product access, go-live and
+ changes all need approval by someone other than the requester, sometimes in
+ several stages. One shared maker-checker component instead of an approval
+ flow per screen.
+- **Provisioning that survives a failing gateway.** Changes to the gateway go
+ through an outbox with retries, and a nightly job reconciles what the portal
+ thinks is live with what the gateway actually has.
+- **Protect the core.** Calls to the core banking system go through a circuit
+ breaker and a cache, so a slow core doesn't take the portal down.
+- **Scheduled jobs that run once.** A single-run scheduler with idempotent jobs,
+ so a job never runs twice when the portal runs on more than one node.
+
+## What the design review found
+
+The review was about questions the design left open, the ones that turn into
+incidents later:
+
+- The session scope of the one-time-password step-up was not defined.
+- The sandbox topology was not decided.
+- Where fee collection ends and the bank's systems begin was not documented.
+- The non-functional requirements had no numbers.
+- Reports would run against the live database.
+
+Each finding came with a recommendation, so the review could be acted on, not
+just read.
diff --git a/content/work/location-tracking-platform.mdx b/content/work/location-tracking-platform.mdx
new file mode 100644
index 0000000..c3e6b79
--- /dev/null
+++ b/content/work/location-tracking-platform.mdx
@@ -0,0 +1,58 @@
+---
+title: "Location tracking for a mobile operator's business customers"
+summary: "The architecture for a platform that lets companies locate their employees' lines and get alerts at geofences, running entirely inside the operator's data centre."
+sector: "Telecom"
+role: "Solution architect"
+themes: ["architecture-integration", "workflow-orchestration"]
+kind: "solution-design"
+period: "2026"
+date: "2026-08-25"
+stack: [".NET", "SignalR", "YARP", "Kafka", "PostgreSQL", "PostGIS", "Redis", "MapLibre", "OpenStreetMap"]
+---
+
+## Context
+
+A mobile operator wanted to offer its business customers a service to locate
+the lines of their field staff, draw areas on a map, get an alert when a line
+enters or leaves one, and report on all of it. Every company sees only its own
+lines; the operator manages them all from a separate panel. It had to run
+entirely inside the operator's own data centre.
+
+I wrote the solution architecture for the proposal: the analysis, the
+architecture and its decision records, a component guide, a comparison of map
+sources, a working map demo and the resource estimate.
+
+## The constraint that shaped everything
+
+Locations come from the operator's network API, which allows ten queries a
+second. At fifty thousand lines, one full sweep takes about 83 minutes, and
+adding more servers doesn't make it faster. So the design had to separate the
+work that waits on that API from everything else.
+
+
+
+## The decisions
+
+- **Workers split by what slows them down.** Collection is limited by the
+ upstream API, processing by CPU, batch jobs by time. Six workers ship as one
+ image, in three deployment groups, so the nightly report never delays a
+ geofence alarm.
+- **A self-hosted map stack.** Map data is processed once a month into a single
+ tile file that the live system only reads. No fee per tile, per user or per
+ request, and nothing leaves the data centre.
+- **Tenant isolation in five places.** Sign-in, gateway, API query filters,
+ row-level security in the database, and separate channels for live updates.
+ If the application forgets a filter, the database still refuses.
+- **Two sites, active and passive.**
+- **Reuse what the operator already runs.** Moving the event backbone onto the
+ operator's central platform removed a whole cluster from the hardware
+ request.
+
+## What I learned from it
+
+The numbers decided the architecture before any technology did. Once the rate
+limit was on the table, splitting the workers, sizing the hardware and
+promising alert times all followed from it.