Biomass burning emissions are a major source of atmospheric compositions of aerosols and gaseous tracers, with important impacts on air quality from regional to global scales. A critical challenge is forecasting wildfire emissions weeks in advance and incorporating them dynamically into coupled prediction systems. To meet this need, we developed a machine learning (ML)-based Wildfire Emission Prediction System using the Light Gradient Boosting Machine (LightGBM) algorithm to predict fire emissions at lead times of approximately 7-15 days. The system runs daily in real time to predict future fire emissions driven by the Blended Global Biomass Burning Emissions Product (GBBEPx), predicted GFS meteorological conditions, and other relevant predictors. The ML-based system predicted FRP is then converted into biomass burning emissions using GBBEPx-derived emission factors together with custom emission factor datasets based on Andreae (2019), VFEI, and RRFS-SD. We found that a six-member lagged ensemble mean of the AI predictions performs better than individual members, providing a more robust basis for emissions forecasting. Preliminary retrospective and near real-time experiments using ML-predicted emissions as input to the UFS-Aerosols model indicate improvements in both weather and S2S forecasts. We evaluated the model prediction performance against MERRA-2, VIIRS and AERONET aerosol optical depth products, as well as surface PM2.5 observations from OpenAQ ground stations. These results highlight the potential of integrating ML-based fire emission predictions into operational coupled models to enhance both air quality and meteorological forecasts across multiple timescales.