Sequence variants are generated using mutation probabilities taken directly from ESM logits, which encode evolutionary and structural constraints and therefore avoid the destabilizing behavior of uniform random mutations. Each candidate is embedded using ESM-C 600M to obtain efficient but information-rich representations for surrogate modelling within a Bayesian Optimization with Multiple Objectives (BOMO) framework, enabling rapid fitness approximation ahead of structural evaluation. Candidate quality is assessed using orthogonal metrics: AlphaFold interface scores (ipTM, ipSAE) capture predicted geometry and interface engagement, while Rosetta dG and shape complementarity (SC) provide energetic and physical validation that reduce overconfident or hallucinated interface predictions. BOMO selects Pareto-optimal designs that jointly improve these complementary metrics. EFNB2 is chosen as the scaffold due to its natural binding to Nipah virus G, but the initial sequence is taken from a specificity-engineered variant to avoid strong off-target Eph-receptor binding in the wild type. This integrated combination of ESM-logit mutation generation, ESM-C-based surrogate modelling, and multi-objective AF/Rosetta evaluation yields EFNB2-derived binders with improved predicted affinity, specificity, and structural reliability.