Technical notes from my work. The list also includes pieces published elsewhere, labelled by publication.
Generative AI is turning into the new Excel, and the labs want the enterprise contracts that come with it. Anthropic and OpenAI are building partner networks the way Microsoft once did, Mistral sells control and sovereignty, and the banks and public sector across APAC leave room for local implementers.
Rules, trained models, LLMs and decision models like Jev each fit a different kind of classification problem. How to choose between them, why combining them often works best, and why any new classifier should run alongside the existing process before you rely on it.
The talk used to be about agents taking work from humans; now it is about fitting them into human workflows. Routine work runs on its own, and the decisions that need judgment or a policy exception reach a person as a prepared case, not an endless approval dialog.
Vector search now ships across the data stack, so semantic search no longer implies a dedicated store and a sync pipeline. The question is what your existing database can do with nearest neighbours in the same query, and whether recall holds up under real filters.
Six tiers, from stuffing the whole corpus into the prompt to a distributed cluster, and what actually forces the move to the next one — in the middle of the range, ranking quality binds before size does.
Intro I use Claude Code and Codex every day. For some types of work, I can stay in the flow by simply reading the output. But I've noticed that agents with…
A tour of the agent frameworks worth knowing — LangGraph and Pydantic AI on the cloud-neutral side, the hyperscaler SDKs on the other — built around one support-triage agent implemented on each.
The laws and standards an AI team actually runs into — privacy law, SOC 2 and ISO 27001, the AI governance frameworks — and which ones to take seriously first.
Cursor, Windsurf, Kiro and the rest all start from the same editor. What separates them is agent behaviour, data handling, and how stable the vendor looks.
Handing out coding assistants is not an adoption plan. A staged path from a legible codebase through specs an agent can act on, to review gates that hold up.
Small, owned, drawn from real traffic, and weighted toward the failures — how to build an eval set that catches regressions instead of quietly rotting.
What AWS, Azure, and Google Cloud actually offer for AI-assisted migration of enterprise code, and where each one fits.
There are plenty of practices which were tried in different types of projects and came from the vast experience of engineers working on them. Examine your…
A personal site does not need much technology. Mine needed to display a couple of paragraphs, a picture or two, and not break. My first idea was therefore to…