28 January 2027 09:00 - 09:30
Your model Is fine. Your plumbing is lying to you
Most AI failures in production come from the infrastructure underneath, not the model.
This talk walks through real incidents caused by data drift, stale features and silent pipeline failures.
In each case the model kept returning confident answers, the dashboards stayed green, and the business paid for wrong decisions long before anyone found the cause. These failures are hard to catch because traditional monitoring checks whether systems are running, not whether the data flowing through them can still be trusted.
The session lays out a practical playbook for infrastructure that doesnât flinch: data contracts at every handoff, feature freshness SLAs, lineage-aware observability, and automated circuit breakers that stop bad data before it reaches a model.
It also shows how to change the question from âis the model accurate?â to âis the pipeline honest?â, and why that one change cuts both incident volume and time to recovery.
Attendees will leave with a diagnostic checklist to find the hidden failure points in their own AI stack, so they can ship faster, break less and reduce downtime.