Attribute
This MVP evolved into a revenue-generating, award-winning AI-powered product used by radio stations and advertisers to evaluate ad performance. It was built on four machine learning systems: predictive attribution, anomaly detection, forecasting, and geofencing. I designed the AI/human interface between those models and the people who had to trust them.
Core challengeAd performance data was imperfect and constantly evolving as the AI's predictions got smarter in real time, making it hard to compare placements or confidently invest ad spend.
Key decisions
Handling data integrity
The AI's predictions shifted as new data came in. I flagged unsettled numbers as "still learning" so users didn't mistake an early machine-generated estimate for a final answer, cutting misreads and building trust in the AI.
Designing for comparison
Rather than spotlight every anomaly the AI detected, I built side-by-side comparison views, so users could spot real trends instead of chasing every automated alert.
Supporting user judgment
The AI could have auto-optimized ad spend on its own. I kept it advisory instead, surfacing AI-generated attribution, forecasts, and geofenced traffic data as context, not a verdict. Humans stayed in control, AI stayed the assistant.
Making time flexible
Users could reslice time windows without losing the "still learning" flags that kept the AI's unsettled predictions from being mistaken for final results.
Annotations
- 1
Traffic, visits, and new users are split out, separating organic activity from what the AI actually credited to the ad.
- 2
Shaded regions mark AI predictions still in progress, real-time machine learning at work, not final truth.
- 3
Upcoming spots use the AI's forecasting engine to project results from past attribution and geofenced traffic.
- 4
Complete vs. ongoing status shows which airings the AI has fully analyzed vs. which it's still learning from.