I have daily WRF model runs that need an automated, production-ready bias-correction layer focused on temperature. You will receive both METAR and SYNOP observations alongside several years of corresponding WRF output; your task is to turn that archive into an AI engine that improves tomorrow’s forecast before it is issued. I am specifically interested in comparing three modern approaches: • LSTM / ConvLSTM for sequence learning • Random Forest or XGBoost as tree-based baselines • Neural-network methods in a classic MOS style Feel free to blend them into a hybrid solution if that raises skill scores. During development please keep RMSE, MAE and Correlation front-of-mind; those metrics will decide which model goes operational. The final hand-off must include clean, well-documented Python code that ingests the latest WRF output, pulls the newest observations, applies the trained correction, and can be scheduled to run automatically every 24 h (cron-friendly). Modular design, clear environment requirements and a short README are expected so I can drop the pipeline straight onto our server. If you have a track record with WRF post-processing, time-series ML/DL and can prove the gains with the stated metrics, I’m ready to start right away and iterate quickly until the bias is tamed.