Saideep Navakoti
I've built banking pipelines for 10 years. My company just told me we're deploying AI agents. I started researching what changes — and found 5 things my pipeline has no answer for:
1. "Build passed" — but the model wasn't tested against this week's data
2. "Health check passed" — but "up" doesn't mean "correct" for AI
3. "Deployment successful" — but which version: the code, the model, or the context?
4. "Cost normal" — but token costs don't appear in your cloud bill like compute does
5. "No alerts" — but the AI failed silently because nobody defined what "wrong" looks like
5 failure modes that your current pipeline has no answer for. 5 slides each. 5 minutes. Drawn from production failures documented in public post-mortems and infrastructure gap analysis from 10 years of banking deployments.
