Keelridge Bank, a fictional composite $46 billion regional bank, has promised investors $42 million in AI savings and $150 million in new personal-loan originations. Release decisions are due in March and OCC examinations of model risk management and BSA/AML start in May. You work through the program as the people who own it: the credit modeler whose challenger model looks good until a proxy explains its lift, the model risk head with 31 overdue validations, the BSA Officer with 7,900 aged alerts, the fraud strategist with a 900-hold-a-day desk and a Friday scam wave, the credit administrator whose drafter put a wrong ratio in front of committee, and the digital banking head whose assistant has to recognize an error notice in a customer's own words.
Each of the six modules solves one business problem with the approaches professionals actually choose between (refresh, build or buy; in-house testing or a platform; rules, ML scores or agents; vendor document AI or GSE tools; keyword routers or gated generative assistants), and every lesson works a full example on realistic synthetic data with the real code, prompt or policy text, its output and an expert critique. You see each approach fail the way it fails in banks (train/serve skew, a reason table keyed by position, labels that lie, an invoice at 2.6 times forecast, a doom loop) and practice the fix in a faded exercise with a model answer. Python notebooks can be run or worked from the printed results, so non-coders are never blocked.
The case does not flatter AI. By the end, the evidence supports less than a third of the promised $42 million as savings the bank can book, net of what the tools cost to run, and about $120 million of the $150 million in originations, and a draft board slide showing $23.7 million has to be taken apart line by line. Learning to say that plainly, with an evidence-gated path to close part of the gap, is a core outcome. Every module ends in a Decision Lab: a branching scenario with business metrics, a cliffhanger event and several endings with an expert debrief. The course closes with a strategic capstone simulation and a release decision package for one use case, written for a model risk committee with a one-page board summary; Ellery gives rubric feedback on it as practice. Real-world examples are cited sidebars attributed to their sources; regulatory and product status is stated as of October 2026, with no endorsements and no vendor performance claims. The course is educational, not legal advice, and the certificate attests to completion and assessment only; it is not a regulatory or professional credential.