ShieldAd

Brand safety violations were costing ShieldAd's clients an average of $1.2M per quarter in wasted spend and reputational damage. Their rule-based system only analyzed page text, missing unsafe visual content like violence in embedded videos or inappropriate imagery adjacent to ad placements. Major clients threatened to pull spend unless detection rates improved above 90%.

Mixpeek's multimodal brand safety pipeline analyzes the full context of an ad placement: page text, surrounding images, embedded video content, and even audio sentiment. A custom taxonomy of 180 brand safety categories was trained with client-specific sensitivity profiles. The system returns a safety score within the programmatic bid window, enabling real-time blocking of unsafe placements.

Integration with ShieldAd's real-time bidding pipeline required sub-50ms latency for safety scoring. The team deployed a dedicated Ray Serve cluster for inference, with pre-computed embeddings cached in Qdrant for known publisher pages. New pages are scored on first encounter and cached for subsequent bids. The rollout covered display inventory first, then expanded to video and CTV over eight weeks.