从元数据到代理：利用 Coactive AI + Databricks 构建内容理解的未来.pdf

1、From Metadata to AgentsBuilding the future of content understanding with Coactive AI+DatabricksAugusto Moreno William Gaviria RojasFlorent BlachotFlorent BlachotVP of Data Science&Engineering7 C F F ;PC L C;2I D;M:b b:Fandoms Fandoms BusinessBusinessCaseCaseFANDOMWorlds LargestWorlds Largestfan plat

2、form 350M+350M+unique visitors per month 250,000+250,000+wikis hosted#1 Source#1 Source for in-depth information on pop culture,gaming,TV and film2.2 M2.2 M500500Imagesuploaded monthlyHours/weekfor manual moderation6?L S(C ABMental health impactfrom exposure to offensive contentChallenge for Trust a

3、nd SafetyChallenge for Trust and SafetyDomain Specific:Domain Specific:Klingon or Blood?Klingon or Blood?Nuances by CommunitiesNuances by CommunitiesAutomatedDomain-SpecificScalable3I F ON C I H I;F MRemove the bad contents and limit exposure for users and partners.Learn nuanced guidelines and apply

4、 these to Fandoms unique content.Handle the volume of Fandom in a cost effective way.5 Evaluation5 Evaluation Points for AI SolutionPoints for AI SolutionPrecisionApprove or reject high percentage of images with no bad images approvedSLAsWork on 99%of images in a few secondsCostBreak even compared t

5、o the current processDev TimeCan be deployed in less than a quarterEvolutiveCan handle use case beyond user safetyFandom Fandom+CoactiveCoactivePOC Part 1(1 mo)Asynchronous TestingFandom curated 20 million images,with PC?I MImagesImagesUsers upload content to Fandom.Coactive ingests content using mu

6、ltimodal foundation models.1212VPCVPC-OF N C GI;F#I HN?HN -I?L;N C I HCommunity Website and ApplicationsCommunity Website and ApplicationsPassionate fansPassionate fansMalicious userMalicious userCoactive selects best multimodal foundation model(s)Coactive selects best multimodal foundation model(s)