Nanobodies are attractive binders for therapeutics and diagnostics, yet predicting nanobody–antigen interactions remains challenging because binding depends on both sequence context and fine-grained 3D structural chemistry, particularly within CDR loops. We present NanoAAI, an interpretable sequence–structure framework that combines protein language model embeddings (ESM2) with atom-level structural graphs encoded by message passing neural networks for nanobody and antigen. NanoAAI integrates these modalities to predict binding while providing attention- and pooling-based attribution signals that highlight structural regions and atomic neighborhoods most responsible for the prediction, enabling CDR-focused interpretation. The model is trained with a supervised interaction objective augmented by contrastive alignment between sequence- and structure-derived representations to improve robustness. Using clustered data splits designed to reduce antigen leakage, NanoAAI improves interaction ranking performance and yields interpretable structural hotspots consistent with binding-relevant regions, supporting practical prioritization of candidate nanobody binders.