Architectural decisions, patterns, and lessons learned from shipping enterprise software.
August 9, 2026
A prompt cannot stop a model from inventing a number
On 1 August our own ad generator produced four different percentages for the same claim, plus a reference to a customer that does not exist. The instruction not to invent figures was already in the prompt. What replaced it: a record of what may be said, a deterministic check that runs before the image is ever generated, and a second check on the image that is itself a model — including where each one stops, and the day the checker hallucinated.
Citations that can be wrong: grounding a RAG answer so the claim and the source actually match
A citation next to a sentence is a claim about that sentence, and four different things can be wrong with it. The failure taxonomy, why a 0.95 faithfulness score can sit on a wrong answer, why LLM-as-judge evaluation cannot be trusted to grade itself, and what to build instead — including the refusal path most systems never ship.
Air-gapped AI: what zero egress actually costs, and what it does not buy
Air-gapped is an operating commitment, not a network setting. The eleven things a normal inference stack phones home about, how to prove zero egress instead of declaring it, the update clock you inherit the day you unplug, and the obligations an air gap does not remove.
Answer once, honestly: the AI security questionnaire in 2026 and the pre-answered pack that survives it
SIG and CAIQ both grew AI sections in 2026, and a SaaS answer set no longer covers them. What actually changed, why the five-actor split is where most questionnaires stall, the ten questions a generic answer library cannot answer, and how to build a pre-answered pack that does not quietly become a set of claims nobody re-checks.
You are the deployer: Article 26 read as an engineering backlog
Almost every AI Act guide is written for providers. If you buy and run someone else's system, your duties sit in Article 26 — human oversight with real authority, input data you control, six months of logs, and telling the people affected. Plus the three ways a deployer accidentally becomes a provider and inherits the whole obligation set.
The AI Act deadline that moved, and the one that didn't: what applies from 2 August 2026
The Digital Omnibus pushed Annex III high-risk obligations to December 2027 — and left Article 50 transparency in force on 2 August 2026 exactly as scheduled. Which of your systems just came into scope, which marking duty lands on 2 December 2026, and what the extra sixteen months are actually for.
Private LLM inference for regulated EU teams: self-host, EU-region API, or US-hosted?
A decision-first, tradeoffs-honest guide for engineering and platform leads: when to self-host private LLM inference, when an EU-region API is enough, and when US-hosted is fine — scored by data sensitivity, the AI Act, DORA and GDPR, latency, and real GPU cost — plus the reference stack and the one design decision that keeps every option open.
DORA for AI systems: a build checklist for engineering and risk teams
A practitioner's build-checklist for bringing an AI system under DORA: each of the five pillars read as an engineering task — what to write down and what evidence to produce, from the Register of Information to concentration risk and the sovereign inference path that reduces it.
Multi-tenant RAG isolation: pushing the tenant boundary into the storage engines
A practitioner's guide to enforcing tenant boundaries in a RAG platform below the application layer — PostgreSQL Row-Level Security with the four bypasses that quietly undo it, per-tenant vector collections and what they actually cost, a test that can fail, and an honest list of what none of it protects against.