Cities across rapidly expanding economies generate staggering quantities of Construction and Demolition Waste (C&DW). Local civil infrastructure cannot keep pace. On smaller building sites, data-driven material forecasting is almost non- existent. Contractors rely on personal intuition. They order raw supplies using rule-of-thumb allowances, causing two distinct on-site operational failures: sudden inventory stockouts that paralyze labor crews, or bloated material piles that spoil into unrecoverable trash. Standard machine learning regressors exacerbate this problem because they optimize for symmetric Mean Squared Error (MSE). MSE treats supply shortages and material surpluses as identical penalties. Real job sites break this symmetry. Running out of concrete during a active pour causes severe financial loss, whereas holding minor excess material carries nominal storage cost. To resolve this structural imbalance, we introduce the Multi-Branch Domain-Aware Waste Estimator (MB-DAWE). Our framework fuses a customized gradient-boosted decision tree (CGBT), a residual subspace ensemble (RSE), and a feature-cross neural network (FCN) through a convex meta-learner trained under a custom Construction-Aware Asymmetric Loss (CAAL). The operational architecture incorporates an on-device Llama-3.2 instance to extract structured data from unstructured field logs, provides instance-level SHAP attributions, and executes a Geospatial Waste Diversion Engine (GWDE) linking excess materials to local recycling plants. Tested on 347,230 CTGAN-generatedrecords matching empirical statistics from 15 Indian cities, MB-DAWE achieves R2=0.8985, RMSE=1.595%, MAE=1.262% and records a Cost-Aware Risk Metric of CARM=80.8 (INR 8,076 per 100-tonne order)—delivering a 10.7% penalty reduction over standard XGBoost.