AdRiver About
In-house Adserver
  • Object oriented model of traffic management for real time decision making
  • Display, Video, Mobile Web, Mobile In-App, Native, Native Video, SmartTV
  • Pretargeting configuration of traffic flows
  • Can offer ad
  • Advanced post campaign reporting and analytics
  • Each ad request triggers auction with both internal and external bidders SSP
In-house DMP
  • Proprietary taxonomy of segments
  • Segments can be used for targeting in DSP solution as well as pushed to customer
  • Target segments are formed based on both own and client data
  • Real time profiling of third party data
  • Cluster analysis is performed on elements of taxonomy first party data to reveal marketing insights
  • Look alike segments generation
AdExchange / SSP
  • Setup configurations ranging from direct orders to multilevel auction
  • Header Bidding is supported
  • Most of RTB protocols are supported for communication with bidders
  • Bid requests can be enriched by first or third party data
DSP / bidder
  • ML optimisation towards conversion of traffic
  • Bidding strategies for 1-st and 2-nd price auction, direct deals, PMP
  • Operations of advertising campaigns are fully automated
  • Bid rate is calculated in real time based on conversion predictor
  • Supports Display, Video, Mobile Web, Mobile In-App, Native, Native Video, SmartTV
  • 2019Q4 average bid-rate =9.7%, win-rate = 11,2% (400B bidreq served)
  • Embedded real-time decision trees
  • Viewability specification — IAB/MRC
  • no MRC certification
  • Fraud detection: SIVT & GIVT
  • Integrated with MRC certified verifiers — DoubleVerify and Adloox Fraud detection: SIVT & GIVT
Infrastructure, architecture, performance
  • 80 bare metal servers, 120K qps scalable to 300K qps.
  • C++, proprietary components from frontend to backend, in-memory databases.
  • System architecture supports working from multiple data centers, allowing for geographically distributed system to serve local users
ML (AI) infrastructure
  • Snapshot of each event consists of 1000 features out of 100K total available features
  • Trained model is used in real time and needs <40ms per bid request
  • Models are trained on data from last 4B events
  • Each bid request results in XXX calculations of primitives on average
  • Architecture supports any ML methods, including hybrids (ie decision trees, logistic regression, naive bayes, etc)
International brands
About Company