Chagas disease detection from electrocardiograms (ECGs) is challenging due to subtle waveform changes hidden within complex signal patterns, variability in ECG morphology, and inherent noise, which demand robust automated diagnostic methods. For the 2025 George B. Moody PhysioNet Challenge, our team Chagas_UBT built an open-source ECG-based system to screen for Chagas disease from 12-lead electrocardiograms (ECGs). Our approach combines a convolutional neural network (CNN), bidirectional long short-term memory (BiLSTM) layers, and an attention layer. To address class imbalance, we use focal loss, ECG-specific data augmentation, mixup, and weighted sampling. The end-to-end pipeline follows the official training interface and returns calibrated probabilities for ranking. We did not perform K-fold cross-validation on the public training data; instead, before training, we excluded 200 validated ECGs for a local holdout (100 positives and 100 negatives) and were kept strictly out of training/model selection and used only for local testing. On this holdout, we observed our best model achieved the score a 0.2941, AUROC = 1.0000, AUPRC = 1.0000, and F1 = 0.82. On the official hidden validation set, our model achieved the Challenge score 0.343. This work points toward earlier detection of Chagas disease, and the open-source implementation with low computational cost makes it practical for screening in resource-limited settings.