Notes from the work: AI, ontology, healthcare data and technical leadership.
Practical writing from real architecture problems. No hype, just what I've seen work, what doesn't, and why.
Six months of AI, platform and data, in one place.
Architecture decision records: the habit I bring to every engagement
The most expensive question in any engineering team is "why did we do it this way?". Architecture decision records answer it in two minutes instead of two meetings. Here is how I use them.
Explainer · AI architectureLLM observability with OpenTelemetry: what to trace, what to leave out, and why it matters
You can't manage cost, latency or quality in an LLM system you can't see. OpenTelemetry now has conventions for exactly this. Here is how I use them without leaking sensitive data into your logs.
Explainer · Fintech · E-commerceIdempotency for payments and orders: how to never charge a customer twice
Networks fail, clients retry and messages arrive twice. In payments and order systems that is not an edge case, it is Tuesday. This is the pattern set I use to make retries safe.
Playbook · Cloud & platformKubernetes FinOps in 2026: scale to zero, resize in place, and know who pays
Two recent Kubernetes releases quietly changed the cost conversation. Here is the playbook I use to turn them into a lower bill, without trading away reliability.
Explainer · Data platformsHybrid search for RAG in MongoDB: when to use $rankFusion, and when $scoreFusion
Pure vector search is great at meaning and poor at exact terms like product codes and drug names. Hybrid search fixes that, and MongoDB now does it natively. Here is how I set it up for retrieval-augmented generation.
Comparison · UAE · RegulationUAE PDPL and AI in 2026: designing for a law whose detailed rules aren't written yet
The federal data protection law has applied since 2022, the free zones run their own regimes, and DIFC already has rules aimed at AI systems. Here is how I turn that patchwork into one platform design.
Analysis · September 2026September 2026 briefing: GPUs get smarter scheduling, idle workloads scale to zero, and observability grows up
The Kubernetes v1.37 deep dives landed this month, and most of them were about running AI and batch workloads more cheaply. OpenTelemetry published evidence that its Prometheus story is improving. Here is what it means in practice.
Analysis · August 2026August 2026 briefing: Kubernetes 1.37, the new MCP roadmap and AI transparency rules go live
A month where workload identity, agent identity and AI transparency all moved at once. Here is what shipped, what it means for production systems and what I would do about it.
Analysis · July 2026July 2026 briefing: MCP goes stateless, the AI Act moves its deadlines, and the UAE gets easier access to AI chips
July was dense. The biggest MCP revision since launch shipped, the EU pushed back its high-risk AI deadlines, US export rules opened up for approved UAE entities, and observability got simpler to switch on.
Analysis · June 2026June 2026 briefing: agents get enterprise identity, telemetry gets columnar, and the FDA codifies change control
June was the month MCP started to look like enterprise infrastructure. Identity moved to the IdP, the SDKs began the switch to a stateless protocol, and in healthcare the FDA wrote change control for imaging AI into regulation.
Analysis · May 2026May 2026 briefing: MCP goes stateless, OpenTelemetry graduates, and the EU agrees to move the AI Act deadlines
The month the agent stack started to look like ordinary infrastructure: a stateless protocol behind a plain load balancer, standard telemetry for LLM calls, and a regulatory timetable you can actually plan against.
Analysis · April 2026April 2026 briefing: Kubernetes 1.36 hardens the platform, open models get a real licence, and FHIR reaches prior authorisation
A month where the boring, important things moved: least-privilege access on every node, an Apache-licensed frontier open model, and a US proposal that puts FHIR into prior authorisation.
Checklist · AI architectureSecuring MCP servers for enterprise AI agents: an architect's checklist
MCP makes it easy to give agents real tools. That is exactly why the security model needs as much attention as the prompts. This is the checklist I work through before an MCP server touches production data.
Analysis · AI regulationThe EU AI Act after the Digital Omnibus: what actually changes for your AI roadmap
High-risk deadlines moved, transparency duties did not. Here is how I read the change as an architect, and what I would and wouldn't put on hold.
Field note · Private AIRunning private LLMs on your own hardware: notes from the bench
I have been running open-weight models on an NVIDIA DGX Spark to understand what local AI can really do for organisations that cannot send data to a public API. These are my working notes.
Playbook · Healthcare dataFHIR to OMOP: a practical architecture for healthcare data
FHIR and OMOP are often presented as competitors. They are not. One is for moving data, the other is for analysing it. Here is how I connect them.
Checklist · AI architectureFrom RAG demo to production: a checklist
The demo took a week. Production is a different job. This is the checklist I work through before an LLM system goes in front of real users.
Explainer · Knowledge architectureOntology vs knowledge graph vs semantic layer: what's the difference, and which do you need?
Three terms that get used interchangeably and shouldn't be. Here is how I explain them to leadership teams, with a worked example and the mistakes I see most often.
Essay · Technical leadershipWhen do you need a fractional CTO? (And when you don't)
A fractional CTO is a good answer to a specific set of problems, and a poor answer to others. Here is how I tell the difference, and what the work looks like once it starts.
Let's talk about what you're building.
The fastest route is a free 30-minute call. Prefer to write? Send a message and I'll reply personally within two business days.