Abstract : Content optimization at internet scale requires decision-making systems that continuously reallocate traffic in response to live reward signals rather than waiting for fixed-horizon experiments to conclude. This article documents the architecture, algorithmic design, and production deployment of a Batched Thompson Sampling (bTS) platform serving 100 million users at 80,000 requests per second across a major internet portal's front page. The system applies the multi-armed bandit (MAB) formulation of reinforcement learning: Beta-Binomial posteriors are maintained over each content arm's expected click-through rate, arms are selected by posterior sampling at each decision point, and parameters are updated on a sub-minute cadence from aggregated streaming reward signals. A dual-pipeline architecture, Apache Storm for near-real-time posterior updates and Hadoop/MapReduce for daily audit-quality reconciliation, closes the reinforcement learning feedback loop at production throughput. Empirical validation on live front-page traffic demonstrated a 3.69% improvement in total clicks over the standard test-rollout baseline, a result subsequently published in a peer-reviewed paper and corresponding to an estimated $100–165 million in incremental annual advertising revenue. The contribution establishes that Bayesian adaptive algorithms are deployable reliably at internet scale and that large-scale production systems can serve as empirical foundations for peer-reviewed research in applied reinforcement learning.