Abstract
Data-centric AI depends on datasets that are diverse, realistic, and structurally reliable, yet real-world data are often scarce, sensitive, or biased. We present DataPiler, an agentic system for context-aware synthetic data generation from natural-language specifications. DataPiler decomposes generation into coordinated stages: query interpretation, schema planning, dependency-aware relational synthesis, domain-context grounding, and iterative validation with hash-based deduplication. The system supports multi-table synthesis with foreign-key consistency and exports SQL, JSON, and CSV formats for downstream analytics. We evaluate DataPiler against Faker, SMOTE, and TVAE on heterogeneous classification benchmarks. On the Adult and Find-a-Buddy datasets, DataPileraugmented training achieves the highest accuracy (0.903 and 0.930, respectively). On a multi-label legal benchmark, context-augmented synthesis improves F1 from 0.4093 (real-only baseline) to 0.5278, while reducing Hamming loss from 0.2646 to 0.1771. These results indicate that an agentic, context-grounded pipeline can improve downstream utility and structural consistency in synthetic data workflows, while leaving formal privacy and fairness guarantees to future trust diagnostics.