Two hours, or four days?
A competitor answered the same application in two hours. Your desk takes four days — and of the seventeen hours a person spends on that file, about one is judgement. The rest is reading, typing, looking things up and copying numbers between systems.
This book rebuilds that day into two and a half hours, of which nearly all is judgement. Not by making the work easier. By moving the reading and the typing to a machine and leaving the part that needs a person — which is harder work than what came before, and the chapter on people says so plainly rather than promising a lighter load.
The part most programmes get backwards
A quarter of that appraisal day is work a plain function can do. Eleven tasks out of twenty-eight: look up a registry, compute a ratio, convert a valuation, apply a formula. No model needed, no risk taken, and the cheapest benefit available.
Start with AI and you skip it. So this book starts with the deterministic platform — nine components, four boundaries that must not be crossed, and five requirements that cannot be retrofitted afterwards at any price. Three of the five can be tested at schema-design time, before a line of code exists.
Then AI arrives, in three widening rounds: one product, one department, the whole bank. Each round ships something that runs.
The invariant core
Take the AI out of the finished system and the process collapses — except for one part, which carries on unchanged. The arithmetic behind a credit limit stays deterministic, because three years from now someone will ask why an application was declined, and the model assessed the risk as high is not an answer.
That shape has a name, and holding it takes four checkpoints around the three agents that read, explain and draft. The third is the one with no equivalent in a general-purpose assistant: every money figure in a draft must be one the engine produced. A fluent sentence proposing a limit nobody computed does not get through.
The code
Eight runnable directories in plain Node.js. Nothing to install, no API key needed, and ninety-four tests that execute offline. Four files are printed in full — a thousand lines — including the agent pipeline and the internal AI gateway.
The tests found three real defects, and all three are described where they belong. One of them was an error in a worked example printed in Book One: an intermediate figure rounded for display and then divided. The book was corrected, not the code.
Who it is for, and what it is not
Architects and engineers will find a specification precise enough to build from. Chief risk officers will find the governance redrawn for the case where a machine decides. Boards will find twenty-two decisions on one page, and the one question to ask Legal in week one.
What this book does not offer: better default prediction in your first eighteen months. That needs your own default data, whatever tool you use. When a proposal promises it, the question to ask is where the training data comes from.
Book Two of two, and written to stand on its own — the opening chapter carries everything from Book One that the rest depends on, and the glossary is repeated in full.
MOBILUCK · code247.ai
