AI in FinTech
About this session
Most large companies now have a Responsible AI policy. It lives in a PDF, it leans on words like fairness and accountability, and almost no engineer has read it. Meanwhile the person shipping an LLM feature has to decide, in code, what the model is allowed to say and do. This talk is about closing the distance between those two documents.
Drawing on work building AI systems inside a regulated bank, I'll follow one principle from the policy statement down to the running system: what it claims on paper, and what it has to become in code before anyone trusts it. Expect real controls you can picture implementing. A vague rule like "no unfair outcomes" has to turn into a test that fails the build when the numbers drift. A model that saw sensitive data in training has to be blocked from repeating it to the wrong person. The interesting part is the mechanics, the actual point where a value becomes a line of code.
Policy and production are usually owned by different people who rarely sit in the same meeting. The space between them is where Responsible AI quietly succeeds or fails. This talk is for anyone on either side of that space who wants to understand how the other half works.
Speaker
Key takeaways
- How to turn one line of AI policy into a control that can block a release.
- Which production failures to guard against first, like data leaks and decisions no one can explain.
- What to log so you have a real answer when a regulator asks why the system behaved as it did.