Optimizing Micro-Targeted Ad Delivery Through Multi-Armed Bandit Algorithms and Neural Networks

Digital advertising suffers from persistent inefficiencies despite a $600+ billion global market, with mistargeted campaigns wasting substantial budgets by optimizing for correlation rather than causal impact. This research presents a novel framework combining neural contextual bandits with causal inference for micro-targeted ad delivery at enterprise scale. We formulate real-time ad selection as a contextual multi-armed bandit problem, deploying deep neural networks for reward prediction while maintaining sub-50ms latency constraints required for programmatic advertising. Three primary contributions are demonstrated: (1) a distributed architecture with hierarchical action space reduction achieving 18× computational speedup, enabling real-time inference across thousands of concurrent advertisements; (2) integration of uplift modeling and doubly robust causal estimation to distinguish incremental conversions from organic user behavior; and (3) adaptive exploration-exploitation strategies balancing performance optimization with advertiser fairness. The AppCapy platform, built on Next.js and Supabase, was evaluated on 8.2 billion impressions across 52 publishers over 90 days. Results show 10-12% improvement in conversion rates versus production baselines, with 28% reduction in cost-per-incremental-conversion through causal optimization. The system achieves 95th percentile latency of 47ms while processing 12,300 requests per second, with near-logarithmic regret growth validating theoretical bandit properties. Causal analysis reveals that 74% of conversions represent incremental lift, demonstrating that traditional correlation-based metrics substantially misattribute advertising effectiveness. This work bridges the gap between bandit theory and production deployment, offering a scalable solution for causally-informed programmatic advertising.