This paper presents a deep learning framework for pricing and hedging American-style options. The method first uses neural networks to approximate continuation values via a variant of the Longstaff-Schwartz algorithm, yielding a candidate optimal stopping strategy. This strategy provides a lower price bound through Monte Carlo simulation. An upper bound is then derived using the dual martingale representation, with nested simulation ensuring unbiasedness. A point estimate and confidence intervals are constructed from the bounds. Finally, the learned continuation values are used to train dynamic hedging strategies by minimizing mean squared replication error, either for the period until the first exercise date or for the entire option lifetime. The approach is tested on Bermudan max-call options under geometric Brownian motion with varying dimensions and parameters. Results show highly accurate price estimates and hedging strategies with near-zero average errors and small shortfalls, demonstrating the method's effectiveness even in high-dimensional settings.