AI automation
reliable and measurable.

A few systems I actually run: a pipeline that drafts, fact-checks and publishes a blog post every couple of days on a site (here is how I build it); a classifier that sorts inbound email and drafts replies in your voice; an agent that scores every incoming lead and routes it to the right person; a quote generator that turns a client brief into a structured proposal. Every time: the machine does the repetitive work, you keep the decision.

Mostly Anthropic Claude (Opus 4.7 for top quality, Sonnet 4.6 for the best price-to-performance ratio, Haiku 4.5 for high-volume tasks). The pick depends on the use case and the budget. If you already run OpenAI, Mistral or Gemini accounts, I can wire those in too.

Three systematic guardrails. First, structured output validated by strict Zod schemas, so the model cannot invent a format. Second, a double pass with web-search fact-checking for anything factual. Third, a human escape hatch: any uncertain output goes to a manual review queue instead of being published automatically. Full write-up on the setup: Automating email triage with Claude.

The Claude API runs roughly $1 to $25 per million tokens depending on the model. In practice, a pipeline that produces 50 articles a month sits around €50 to €80 of Claude usage. The bulk of an automation project is the initial build, not the long-run usage. Model details and pricing are kept current across my blog.

A simple automation (email sorting, template generation): 2 to 4 weeks. A full pipeline with validation, monitoring and a human fallback (like the article generator I run on my own blog): 6 to 8 weeks. The code stays yours, on Git, with no mandatory subscription to my services. Want a website at the same time? My custom websites service picks that up.