In this paper, a stochastic optimization framework is developed to reduce congestion on distribution feeders using batteries, under offline and online design paradigms. Our design is customized, implemented, tested, and analyzed in a real-world testbed that was built based on a university-utility collaboration in California. Our proposed method seeks to optimize peak load at the feeder while taking into account feeder load uncertainty as well as hardware, utility, and customer constraints. We present both experimental and numerical results. Insightful observations, design trade-offs, and lessons learned are discussed.