Notes from the harbor
Practical, occasional writing on AI and data — what works, what does not, and why.
RAG that actually works in production
Most retrieval systems fail not on the model but on the plumbing. What separates a demo from a dependable assistant.
Read →The EU AI Act, for Nordic teams
A practical map of what the Act asks of you, when, and how to build governance in from day one.
Read →Agentic AI: when not to use it
Autonomous agents are powerful and easy to misapply. Where they earn their keep, and where a plain pipeline wins.
Read →Why data quality beats model choice
Teams obsess over model architecture and neglect the data feeding it. The returns are almost always the other way round.
Read →MLOps without a heavyweight platform
You can monitor and retrain models reliably with light, open-source tooling. Here is the minimum that matters.
Read →The semantic layer, explained simply
Why one agreed definition per metric is the single highest-leverage fix for reporting nobody trusts.
Read →Vector search: five common pitfalls
Chunking, embeddings and evaluation trip up most first attempts. A checklist to avoid the usual traps.
Read →Feature stores: when they are worth it
A feature store solves real problems and creates new ones. When the trade actually pays off, and when it does not.
Read →Keeping LLM costs under control
Token bills scale quietly until they surprise you. Practical levers to cut cost without cutting quality.
Read →Data contracts in practice
A data contract is a promise between producer and consumer. How to make them real rather than a diagram.
Read →How to evaluate LLM outputs
Human review does not scale and vibes are not a metric. Building an evaluation harness you can trust.
Privacy by design in AI systems
GDPR is not a checkbox at the end. How to build data minimisation and purpose limitation into an AI system from the start.
Streaming vs batch: choosing honestly
Real-time is fashionable and often unnecessary. A clear-eyed look at when streaming earns its complexity.
Model cards that actually matter
Most model cards are box-ticking. What belongs in one so it is useful to auditors and engineers alike.
When not to use AI at all
Sometimes the honest answer is a rule, a form or a spreadsheet. Knowing when to say no to AI.
The Harbor Log
Occasional, practical notes on AI and data. No noise.