Seekr
Senior Product Designer
Align
Brand safety scoring for podcast advertising
Role
Senior Product Designer
Team
ML dev team, Researcher, VP of Product Management
Partners
Oxford Road
Context
A VP walked up to me with a sketch and an idea. Two weeks later we had a prototype in front of media planners. Three months later we shipped a patented AI product to thousands of users. Along the way, our design research changed how the AI model itself needed to work.
Scanning episode
PROFANITY12:30
Political29:38
12:3012:4813:0113:3514:09
Listening/Reading
Align is reading the episode end-to-end. No brand-safety signal has surfaced yet — the timeline above fills in as each segment is cleared.
Auto-playing the read — watch the timeline resolve
The Challenge
Brands need to know the podcasts they advertise on are safe to their level of comfort. Seekr's AI scores content across 11 GARM risk categories and our proprietary Civility Score, but raw AI scores don't mean anything to a brand manager scanning 500 podcasts.
AI scores alone don't build trust. In fact, as we all know, they are non-deterministic, and they definitely can't sense sarcasm.
A low politically sensitive score that's acceptable for one advertiser is a dealbreaker for another.
I partnered with our researcher Nora to map brand-type thresholds and design a journey that guided users to the right context.
Our unique value
The hardest part wasn't visualizing scores, it was understanding what those scores meant to different brands at each level of exploration.
Our sweet spot in the user journey was helping them say yes or no quickly while scanning thousands of podcasts. Our patented Civility Score displays, measured by number of attacks, is our certified rating of the show. It is based on general and severe attacks found in the show.
General attack - insulting language toward an individual or group.
Severe attack - based on characteristics with legal protections against discrimination
3 levels of context
Brands can drill from high-level podcast scores down to the exact audio moment that triggered a flag. This is the closest we can get to AI explainability in such a short time frame, which in most cases helped the users too! In cases where the model failed, we let users give feeback, thumbs up or thumbs down + open input field. Feedback was only given about 8% of page views but every little bit helps retrain our model.
Podcast Overview
Browse all podcasts with aggregate scores. GARM risk levels and Civility Scores visible at a glance.
Episode Breakdown
See every episode scored individually. Identify which episodes carry the most risk and why.
Flagged Moment
Jump to the exact part of the episode that triggered a score. Audio player anchored to risk context.
