customers-RedNote

Customer Introduction

Red Note is a leading lifestyle-sharing platform and a pioneer in social e-commerce. With hundreds of millions of monthly active users, it empowers people to share life experiences through short videos, photo-text "Notes," and live streaming. As a content-driven community, Red Note maintains a high brand moat regarding authenticity and compliance, representing one of the most challenging content moderation scenarios globally.

Scenario

Red Note business spans all dimensions of social interaction, facing highly complex risk control requirements:
1. Full-Scenario Compliance: Covers Community Notes, Comments, User Profiles, Private/Group Chats, Live Streaming, and Ad Placement, requiring detection of political, pornographic, violent, minor-protection, and anti-fraud risks.
2. Multi-modal Recognition Challenges: Notes and ads contain vast amounts of Image OCR, brand logos (for counterfeit detection), audio "moan" detection, and voiceprint recognition.
3. Ad Compliance & Brand Protection: Ad placements must strictly adhere to advertising laws while identifying sensitive information like counterfeit logos and private license plates.
4. Real-time Live Streaming Needs: Live video and audio require high-frequency, low-latency interception of violations, including specific scenarios such as unauthorized appearances of minors.

Solutions

DeepCleer in collaboration with Xiaohongshu, built a deeply integrated solution combining AI auditing, custom models, and refined enforcement:
1. Deep Multi-modal Fusion: Implemented a combined audit mode where "Image OCR content is merged into text" to ensure semantic consistency; utilized proprietary keyframe extraction in live streams to reduce audit volume by 2/3.
2. Asynchronous RTC Audio Slicing: Employed a non-fixed frequency "silent segment slicing" logic for audio, integrated with ASR transcription, moan detection, and prohibited song libraries for comprehensive audio governance.
3. Advertising Law & Brand Specialization: Developed specialized Brand Logo recognition combined with merchant credential verification to combat counterfeit promotions; implemented automated license plate desensitization for user privacy.
4. Refined Tiered Enforcement: Supports over ten types of enforcement actions, including "Normal Exposure, Self-visible (Author only), Fan-only, Exposure Reduction, Blocked, and Room Closure," balancing community atmosphere with regulatory mandates.
5. Targeted Inspections for High-Risk Labels: Established a DeepCleer sensitive word ranking and periodic cleaning mechanism for N1-level baseline issues (e.g., political, minors) to ensure sustained risk control stability.