Blog
Arguments, use cases, and working notes.
Published here first. Some are arguments; some walk through a single use case end to end. All of them come back to one position: own your context, rent your intelligence, and be able to prove what the system did.
Published
16 so far.
Retrieval is not one thing
"RAG" now covers five meaningfully different architectures with different strengths, costs and failure modes. Choosing between them is an engineering decision — and the most common choice is wrong for enterprise questions.
2026-10-05Hallucination is a pipeline problem
You cannot prompt a model into telling the truth. You can build a pipeline in which an unsupported answer cannot ship. The difference is five enforced gates — and a willingness to refuse.
2026-10-05The model is the cheapest part of the agent
Every agent framework demo shows you the model. Production agents are made of something else — the harness around the model that decides what it can touch, spend and do. That harness is the product.
2026-10-05Your context should not leave the building
Sovereign AI is usually sold as where the server sits. The sharper question is where your context goes — the records, relationships and history that make answers about *your* organisation possible.
2026-10-05The five levers on an AI bill
Enterprise AI bills are not driven by one expensive model call. They are driven by systems that pay to re-read the same context all day. Five levers move most of the spend, and none of them is negotiating the price list.
2026-10-05The agent prepares. The officer decides.
What an approval gate looks like from the officer's side of the desk: the case arrives prepared, the evidence is attached, and the decision — which was always theirs — finally takes minutes instead of days.
2026-10-05The green screen that answered back
Every large estate has one: the system with no API, no vendor, and thirty years of the truth in it. Ignoring it is not an option and replacing it is not a plan. There is a third route, and it is more honest than both.
2026-10-05A wildfire alert with a coordinate attached
A detection that arrives as a coordinate on a map, with the image and the confidence attached, is a different thing from a camera feed someone was supposed to be watching. What it takes — and what it does not do.
2026-10-05Proving the target equals the source
Every platform migration ends at the same uncomfortable moment: someone has to say the new system matches the old. That sentence is usually backed by samples and optimism. It can be backed by a full comparison, on live data, instead.
2026-10-05The invoice that was paid twice
Somewhere in this quarter's payables is an invoice that was paid twice, and another that was billed above the contracted rate. Sample-based audit will not find them. Reading every line will. Here is how the check works.
2026-10-05Ten resolved tickets, checked before anyone looks
A service desk closed ten tickets this morning. An agent re-checked each one against the systems themselves. Seven were genuinely resolved. Three were not. This is what verification actually looks like.
2026-11-29Stop asking whether it runs on premises
Every vendor can say yes to on-premises. The useful questions are whether it is the same product, and whether the AI still works when nothing can leave.
2026-11-15Singapore published a framework for agentic AI. Here is what it asks of you.
A plain reading of IMDA's Model AI Governance Framework for Agentic AI: who it is for, its four dimensions, and what an organisation deploying agents would need to do.
2026-11-01When a knowledge graph is the wrong answer
Graph retrieval is over-applied. Here are the cases where plain vector search wins, where a graph is over-engineering, and the criteria that should decide.
2026-10-18Your AI pilot did not fail for the reason you think
Teams blame the model when the documents were the problem, and the documents when the retrieval was. The symptoms match and the fixes are opposite. Here is how to tell.
2026-10-04A citation is not evidence
A link that resolves proves a source exists, not that it says what the sentence claims. Here is why the gap matters and how to measure it.