Design the AI platform for a 5,000-employee bank

platform, security, evals, knowledge, production, governance

Prompt

You're the staff engineer asked to design the AI platform for a 5,000-employee bank: one platform that multiple business units, retail support, fraud review, internal engineering, and compliance research, will build AI features on top of. Walk through the platform's architecture, from the model and tool layer up through security, evaluation, and governance.

Interviewer's answers to clarifying questionsopen after you have asked your own

Business units and use cases? Retail customer support (chat, policy Q&A, limited account actions), a fraud-review assistant (summarizes cases for human analysts, never decides autonomously), an internal coding agent for engineering, and a compliance research assistant over regulatory documents.
Regulatory context? The bank is subject to strict regulatory obligations, data residency, auditability, model risk management, and EU customers trigger EU AI Act obligations for some features.
Data sensitivity? Customer PII, account data, and internal case files are all in scope for at least one use case; strict tenant/business-unit isolation is required internally.
Current state? Each business unit has been prototyping independently with different vendors, different logging, and no shared security review.
Scale target? The platform must support at least 10 business-unit teams building on it within 18 months, with a shared on-call rotation for platform-level incidents.
Board-level ask? A single governance answer the CISO and Chief Risk Officer can both sign off on, plus a cost ceiling the CFO can hold teams to.

Follow-upsafter your high-level design

The fraud-review assistant reads case notes that include customer-submitted text, an untrusted-input channel, and has access to account data. Walk through the specific risk this creates and how the platform's shared controls address it versus what each team must still do itself.
A single frontier-model provider has a multi-hour outage. Which of your four use cases can degrade gracefully, and which cannot, and why?
The compliance research assistant cites a regulation that was superseded last month. Whose eval/observability layer should have caught that, and how?
18 months in, a new business unit wants to onboard a "super-agent" with access to all four domains at once, to save integration work. What does the platform's design let you say no to, and why?
The CRO asks for one number representing "how safe is our AI estate" for the board. What do you actually give them, and what do you refuse to collapse into one number?

Answer first: out loud, on paper, or with the interview prompt above. Then compare with the rubric.