An already overburdened national power distribution infrastructure faces serious operational issues as a result of Pakistan's quickly expanding electric vehicle (EV) sector. In addition to limiting the efficient use of available renewable energy, uncoordinated commercial EV charging results in severe peak demand spikes, voltage violations, and increased energy expenditures. A decentralized, machine learning-driven load scheduling architecture for Pakistani commercial EV charging stations is proposed in this study. Using a dataset of 35,040 hourly observations from NEPRA, NTDC, PESCO, and Pakistan Meteorological Department sources, a Random Forest Regression (RFR) model is created using the Python scikit-learn library. It incorporates 14 engineered features, such as time-of-day, feeder loading, ambient temperature, solar irradiance, time-of-use (ToU) tariff windows, and lag-based load features. The trained model attains an R2 of 0.946 and a test Mean Absolute Error (MAE) of 4.73 kW. A priority-weighted scheduling algorithm that maximizes charger activation within transformer capacity and NEPRA voltage regulation restrictions is fed the projected demand profile. In comparison to an unplanned baseline, simulation results for a sample 630 kVA urban Peshawar station show a 27.4% decrease in daily peak demand, a 23.1% decrease in daily energy expenditure, and a 19.8 percentage-point increase in on-site solar PV use. The edge-deployable architecture is ideal for Pakistan's limited SCADA and ICT environment because it does not require real-time cloud access.