Conducting Intelligence

Large language models write a great deal of the code now. In a regulated industry, you still have to answer for every line of it.

Conducting Intelligence is a discipline for building real systems with AI where "the model suggested it" is not a defence: medical devices, automotive software, banking and payments, anywhere an auditor, a regulator, or a certification body will eventually ask you to account for what you built.

It rests on two tools. Felicity is an architecture-first engineering framework, forged in safety-critical firmware, that turns vague intent into verifiable requirements, interfaces, and tests before any code is written. The conductor model is a way to orchestrate many AI agents under a single accountable human, with every decision mediated through files, so an audit trail exists by construction. Neither needs the other. Together they make large projects built with AI governable: a human stays accountable, the work is reconstructable end to end, and the system actively defends against the failure modes that put you in front of a review board.

You see it applied, not just described. The book carries one regulated build all the way through, a card-payment switch developed against PCI DSS and card-scheme certification, from first requirement to passing tests. And it maps onto the regimes you already work under, IEC 62304 and ISO 13485 in medical, ISO 26262 in automotive, PCI DSS and model-risk governance in finance, treating traceability, change control, and documentation as first-class rather than afterthoughts.

This is not a book about going faster. AI already made everyone fast. It is about being able to stand behind what you shipped. The promise is quieter, and rarer: less regret.