This paper presents a compact, clinically oriented pipeline for thyroid ultrasound decision support that couples a fine-tuned ResNet50 classifier with BLIP captioning and a LangChain-based multi-agent analysis layer. A curated dataset of ≈2,549 thyroid ultrasound images (normal/abnormal) is processed through a lightweight Streamlit interface that handles upload, preprocessing (resize/normalize), and dynamic model loading. The classifier is trained with Adam and progressive layer unfreezing, and evaluated using K-fold validation with ROC analysis. A BLIP captioner, adapted to the medical domain, generates concise image descriptions that feed a medical analysis agent (for TI-RADS–style reasoning) and a data analytics agent (trend/risk context). Real-time explainability is delivered via Grad-CAM, SHAP, LIME, and edge overlays, and the system exports structured PDF reports summarizing predictions, confidence, captions, and rationale. On the held-out test set, the model attains 99.56% accuracy, 99.20% sensitivity, 99.80% specificity, and AUC = 0.998, demonstrating high discriminative performance while remaining interpretable at the point of care. The contribution lies in unifying robust classification, clinically useful captions, agentic reasoning, and multi-modal explanations within a single deployable workflow, supporting transparent decision making and streamlined documentation for radiology clinics.