Practice 01 · Platform & integration architecture
The seams between systems are where the cost actually lives.
Most platforms are not slow or fragile because a service is badly written. They are slow and fragile because the contracts between services were never stated, so every integration is a negotiation and every change is a surprise.
-
22s → 0.4s
report resolution, by collapsing roughly 1,500 sequential queries into about seven batched
-
285+
device capabilities across 57 vendors, under a 1,279-test conformance suite
-
20+
partner integrations shipped and maintained, not prototyped
What I do here
I make the next integration boring.
I design the substrate that makes it so: canonical data contracts, event streams other teams can build against, and a capability model that turns new vendors into data rather than into a release.
-
Event-driven and event-sourced systems
Kafka, Debezium change data capture, Kubernetes, AWS. The decision that matters is rarely which broker — it is whether the events you publish are a contract other teams can build against, or an implementation detail you will regret exposing.
-
Partner integration fabrics
Twenty-plus integrations, each with a partner who versions on their own schedule and reads the spec differently than you do. The work is canonical data contracts and an anti-corruption layer, so one partner’s idea of a unit does not leak into your domain model.
-
IoT control planes
A capability catalogue that describes what a device can actually do, rather than a switch statement per vendor. That is what makes 57 vendors tractable: new hardware becomes data, not a release.
“One of their standout achievements was taking our multi-vendor integrations service, previously unreliable and difficult to use, and rebuilding it from the ground up into a robust solution that dramatically improved our customer satisfaction.”
On the record
Recent work in this practice, newest first.
-
2026
Five radios on one microcontroller
LoRaWAN, BLE, Z-Wave, NFC and Wi-Fi in a single Zephyr RTOS image with secure boot and signed over-the-air updates — at 73% flash and 27% RAM. The budget was assumed to be exceeded; I disproved that with measured builds rather than argument, which is usually the cheaper way to win an architecture disagreement.
-
2026
A user interface that writes itself
A server-driven UI where the interface is a projection of a machine-readable capability catalogue: add a device capability and its controls appear for entitled users with no renderer changes. One contract drives a web runtime plus generated Android and iOS native code, verified by real compiler runs.
-
2025 – 2026
An event-sourced IoT control plane
Built inside a governed multi-agent delivery programme, alongside a greenfield twelve-service platform and shipping firmware — the integration substrate the device work above depends on.
One of four practices
The architecture is only half of it.
What decides whether a platform can be trusted is the governance layer on top of it, and how the work gets delivered.