At Next Advertising (Nexta) our mission is to bring simplified, automated and optimized online marketing to our enterprise and SMB clients on very large-scale. While only being 2 years young, our mature platform product already serves dozens of enterprise clients, and through those thousands of SMB globally.

From a non-technical point of view, the task of the AI agent is to continuously make adjustments to live campaigns in order to improve relevant business KPIs, typically cost per click or conversion. From a technical point of view, we frame the task as a sequential decision-making process under uncertainty, and employ techniques from reinforcement learning, control theory and machine learning to solve this non-trivial task. While data in online advertising is usually abundant, this is not the case for the short-lived, customised and few-conversion campaigns that our clients typically run.

For this reason, we heavily employ Bayesian techniques, from building multi-layered conversion funnel models, where propagation of uncertainty between the layers is imperative, to balancing the exploitation-exploration trade-off in the final optimization step for making the best decision. Thousands of live campaigns are currently running in the Nexta system, and deploying state-of-the-art AI solutions on this scale, requires the use of modern cloud-based infrastructure solutions. These are developed internally in the AI team, aided by senior developer resources, and developing production-level quality code is part of the job of an AI developer.

Keywords:

• Bayesian hierarchical multi-layered models, multi-task learning, autoML, experimental design, Bayesian optimization, sampling techniques, probability and uncertainty.
• SGD, constrained and risk-aware optimization.
• Python, stan, pytorch, docker, microservices, event-driven architectures