Automated evaluation of open-ended student work remains a challenge in educational technology. In the context of SQL query assessment, existing models often rely on rigid heuristics or underfit architectures that fail to generalize. Here we present a multi-objective neural model whose architecture and hyperparameters are optimized using evolution strategies (ES). Our model jointly predicts query correctness, diagnostic remarks, and numerical grades from raw student submissions. We show that this approach improves classification accuracy and robustness across underrepresented feedback classes, while maintaining interpretability. These findings demonstrate the utility of ES in discovering high-performing configurations for complex assessment tasks.