Writing
Occasional notes from production work
Short write-ups on things I've built and what I learned building them — EU-sovereign AI, embedded DSP, MLOps, and the in-between.
-
Agreement is not verification.
A budget arrived from another session with an €8,000 GPU on it, marked "locked." Two models agreed on the number. Reading the actual spec, and making a different model family check the first, collapsed it to a €750 card. A confident plan and a correct plan look identical.
-
AI review is plausibility review.
In a 68-equation theory document, two AI reviews in assessment mode caught zero errors. A third pass with the same model family but a reproduction instruction caught three. AI review is plausibility review unless you make it not be.
-
Verification is a workflow problem, not a model problem.
A frontier model drafted a verdict on a €5k decision. Two fresh-session reviewers caught fifteen issues with near-zero overlap. The failures lived in the drafting context, where model-level interventions cannot reach.
-
Most production AI grades its own output.
Most production AI lets the LLM grade its own output. Notes from a year of building one that does not, and the second decision that makes the first one honest.
-
Senior developers trust AI less than juniors.
Three 2025 surveys keep finding the same pattern: more experience, less trust in AI output. The data flips the conventional take that AI is a junior-shaped problem.
-
The work is splitting. The conversations haven't caught up.
Engineering work is decomposing into producing output and validating it. Production is now cheap and fast. Validation is not. An observation from a few years of building with AI agents.
-
A small GDPR-safe chatbot
Notes from a side project that stuck around. Observations on location as architecture, why chunking beat embeddings, and how the European AI picture shifted in a year.