Writing
Notes on building AI products
Lessons from taking AI from prototype to production — engineering, architecture, and the decisions that decide whether an AI feature ships.
AI agent maintenance: model changes, regressions, and what a maintenance contract has to coverWhy an AI agent needs maintenance even when nobody touches the code: model deprecations, upstream systems that change formats, knowledge sources going stale, new attack patterns. How regressions get caught, how a model migration runs, and the ten points a maintenance contract has to contain.Operating AI agents: what happens after go-live, and who is responsible for whatAn AI agent is not finished at go-live. What operation actually means: approval mode, logging, evals in production, cost budgets, the incident path, the review cadence, and how the duties under the EU AI Act and the GDPR split between the company and its software vendor.Your customers don't google anymore — they askAI assistants are replacing classic search for buying research. What the 2025–26 data really says about AI Overviews traffic and click-through — and an honest generative engine optimization (GEO) playbook for AI search visibility.Agentic software in the Mittelstand: a practical, GDPR-aware guideAI agents are moving from vendor demos into real Mittelstand operations. What agentic AI actually is, where AI agents create value in mid-sized companies, when a deterministic workflow beats an agent — and how to adopt agents without violating GDPR or the EU AI Act.RAG vs fine-tuning for SMEs: an engineering-economics guideRAG vs fine-tuning is the wrong fight — they solve different problems. What retrieval augmented generation actually costs, when fine-tuning an LLM earns its keep, and a worked cost example for a 10,000-document knowledge base, with current provider pricing.Shipping AI that survives productionWhy most AI features die between demo and production — and the engineering discipline that gets them across: LLM evaluation harnesses, LLM-as-judge calibration, observability and tracing, guardrails, and cost budgets. A field guide to eval-driven development and AI reliability.
What work could agents take off your team?
Bring one recurring workflow or a product idea. Thirty minutes, an honest assessment, a clear next step.