This paper presents a deep reinforcement learning approach to service-level-agreement (SLA) aware network slicing in 5G radio access networks. We formulate the dynamic allocation of Physical Resource Blocks (PRB) among three coexisting slices — enhanced Mobile Broadband (eMBB), Ultra-Reliable Low-Latency Communications (URLLC), and massive Machine-Type Communications (mMTC) — as a Markov Decision Process and develop a Gymnasium-compatible simulation environment in which per-slice SLA requirements follow the 3GPP TS 28.530 framework. A Proximal Policy Optimization (PPO) agent is trained with a composite reward that jointly rewards aggregate throughput, penalizes SLA violations, and encourages resource utilization. The agent is evaluated over 100 episodes against three classical baselines: Static Equal, Static Priority, and Greedy Proportional allocation. The PPO agent reduces the aggregate SLA violation rate to 2 × 10⁻⁵, three to four orders of magnitude lower than the Static Priority baseline (0.46), while sustaining a mean aggregate throughput of 74.7 Mbit/s and a resource utilization of 98.6%. A throughput-versus-violation analysis confirms that the learned policy occupies an operating point that no single static rule attains. The complete environment, training pipeline, and evaluation scripts are open-source and reproducible using the Stable-Baselines3 library.