AI Companionship

Talkie

How a leading AI-native social platform cut false negatives from 4% to 0.2% across text, voice, and NPC imagery.

4% → 0.2%

False Negative Rate

20% → 7%

False Positive Rate

Languages Covered

The Customer

An AI-native community
built on living characters.

Talkie is a leading AI-native content community and social platform where users engage in real-time dialogue, voice calls, and role-play with thousands of AI-driven virtual characters (NPCs).

Its core experience is built on highly autonomous AIGC interactions — AI text responses, intelligent voice calls, and personalized NPC cards created from user prompts.

At a Glance

IndustryAI Companionship

Content ModelAIGC dialogue + UGC NPCs

MarketsEN · ES · PT · JA · ZH

DeepCleer ProductsText · Audio · Image Moderation

The Challenge

Complex risk —
at global scale.

CHALLENGE 01

AIGC Dialogue Risks

AI-generated text and voice outputs can produce compliance risks — sensitive political content, terrorism, NSFW, or abusive language — that emerge dynamically and can't be caught by static filters.

PoliticalTerrorismNSFWAbuse

CHALLENGE 02

NPC Creation Compliance

NPC images generated from user prompts may contain NSFW content, nudity, child-safety violations, or violent imagery, requiring strict alignment with platform guidelines.

NSFWChild SafetyViolence

CHALLENGE 03

Global Multilingual Pressure

Operating across English, Spanish, Portuguese, Japanese, and Chinese, the platform must balance free expression in overseas markets with rigorous interception of P0 high-risk content.

5 LanguagesP0 Interception

The Solution

A closed-loop strategy
tuned to the platform.

Comprehensive AIGC Dialogue Defense

Full-volume, real-time detection for AI-generated text and voice calls. Deep recognition for English, with a "Translation + Policy Audit" model applied to Spanish, Portuguese, and Japanese for global consistency.

NPC Visual Safety Check

Detection is integrated directly into the image generation phase — if the AI audit issues a "REJECT," generation terminates immediately. Passed images proceed to human audit for style judgment and operational tagging.

Differentiated Regulatory Strategy

Tailored for global markets — focusing interception at the model output stage while maintaining flexibility for user inputs to balance compliance with user experience.

High-Efficiency Quality Assurance

A routine inspection mechanism performs large-scale daily spot-checks on facial data and weekly false-positive evaluations to drive continuous policy iteration.