Insights
3 pieces · technical practice, not client specifics
Written from engagements, not from a content calendar.
Scroll into it
Latest · 4 min
Why AI systems fail in production, not in the demo
A model that scores 94% on your eval set can still fail a third of the time in production — usually not because the model got worse, but because the eval set never looked like production traffic.
Read thisMore from the practice
01 / 02
What the EU AI Act actually requires, in plain language
The EU AI Act sorts every AI system into a risk tier, and the tier — not the technology — decides what you're legally required to do. Most companies deploying AI fall into "limited risk" or "high risk" without realizing which one, or what it actually obligates them to.
5 min · 13 sections
Read the piece02 / 02
Retrieval architecture decisions that matter more than model choice
Swapping a language model rarely fixes a retrieval-augmented system that's giving wrong answers. If the top-k results it's grounding on are wrong, a better model just states the wrong answer more fluently.
4 min · 7 sections
Read the pieceWhat these are
These are notes from building the systems, published because the failure modes are worth knowing before you commission one.
They describe general practice — nothing here is drawn from a client engagement.