RAG Reality Check
by Stéphane KREBS
Audits what a RAG system really does, not what its docs or scores claim: the executed code path, silent failures, retrieval measured on its own, and the blind spots of metrics like faithfulness.
I build agent skills for teams who can’t afford an AI that guesses. My path: maths and physics teacher, Java developer, PhD in statistical modelling (Sorbonne), then 300+ professionals trained in change and innovation, and 225+ certification audits. Today I design AI systems on regulated data, and I package the methods that held up as skills. Every skill I publish follows three rules: facts are extracted by code, not invented by a model; the agent’s work is checked by a reviewer that didn’t write it; and no performance figure appears unless it was measured. Questions or a custom need? Reach out: I answer.I build agent skills for teams who can’t afford an AI that guesses.
by Stéphane KREBS
Audits what a RAG system really does, not what its docs or scores claim: the executed code path, silent failures, retrieval measured on its own, and the blind spots of metrics like faithfulness.
by Stéphane KREBS
Run a coding agent like a careful team: written spec, one reviewed batch per branch, independent reviewer subagents, a stop hook that blocks secrets, and reports with measured numbers only.