We present a production hierarchical Bayesian framework for estimating heterogeneous treatment effects in large-scale online advertising experiments. By jointly modeling campaign-level effects and cross-campaign heterogeneity, it addresses a limitation of inverse-variance weighting, the standard meta-analytic baseline, which becomes unstable in such challenging regimes. Across simulations spanning a range of effect sizes and heterogeneity levels, the approach delivers more reliable inference under high heterogeneity while matching or improving estimation accuracy overall. We further demonstrate its practical value in a production demand-side platform handling over 5 million requests per second, integrating it with sequential testing and evaluating it across three experiments covering over 35,000 campaigns from 3,000 advertisers.