How to warn developers without scaring them?
Designing an alert that detects overfitting and suggests ML training improvements.
"My model is 100% accurate. Why is it failing in production?"
You are a citizen developer, using Appian's AI skills. You confidently train extraction models to near-perfect accuracy, then watch them fail in production.
The root cause is overfitting: models trained on small datasets memorize patterns instead of "learning" from training data, leading to high training accuracy but low production accuracy.
The hardest part wasn't visual design…
it was designing a unified experience, crafting language and visual cues with intention.
My job was to surface the overfitting problem at the right moment, in language developers could act on, without shaking their confidence in the platform. During the design process, I identified a mismatch between the alert's urgent visual design and its educational copy, and drove a cross-functional revision with my team to resolve it.
The feature shipped in Appian 23.4. The team estimates a 20% reduction in data labeling errors across enterprise customers.