REGU

REGU started as a RAG system that ingested the EU AI Act, effective August 2026, and produced structured, cited compliance reports for founders asking which obligations apply to them. Law, guidance, commentary and examples lived in four vector tables with their own chunking; hybrid search blended semantic and lexical scoring, and a confidence gate failed closed when citations did not meet a threshold. Every claim traced back to a chunk ID.

Conversations with the founders it was built for moved the problem. The question they actually had was about cost and control: what would it take to run their models on their own hardware, and when does that pay off. REGU became that answer: a local LLM infrastructure setup and maintenance service for startups and companies in Munich.

The first product of the second chapter is a one-page analysis report. A company enters its current cloud AI spend and usage profile; the engine picks the most suitable local hardware from a catalogue and shows what they pay today, what they would pay with REGU, break-even, ROI, a three-year TCO and an urgency score. The calculation engine is locked by a golden output shared with a Python prototype, so a formula change is a deliberate, reviewed diff.