The Winners Decided Where Not to Use AI

VitaSort reorders an outpatient queue by urgency but refuses to let a generative model make that call, using the deterministic NEWS2 score so behaviour stays predictable. Toddle AI runs 33-landmark pose estimation and then applies deterministic gait analysis, and states plainly that it is a research prototype rather than a diagnostic. DocSync opens a merge request when confident and hands a human an issue when not. The OpenMRS consultation stack builds in the assumption that a clinician reviews the output first.

Each of these designs where to stop rather than how far to automate. Hackathon scoring tends to drift toward degree of automation, so the opposite instinct standing out tells you judges are imagining real operation. In domains like healthcare or HR, where being wrong is expensive, a project with no defined stopping point simply cannot be deployed.

Takeaway: identify the single place in your build where you least want to be wrong, and replace just that with deterministic logic or a human check. Then explain the decision explicitly in the demo. 'We deliberately don't let the model decide this' reads as understanding the domain, not as a lack of technical nerve.