AppCompiler: A Multi-Stage LLM Pipeline for Generating Validated Application Schemas from Natural Language

Generating application architectures from natural language descriptions is a challenging problem requiring cross-layer consistency across database schemas, API definitions, user interface specifications, and authentication configurations. Existing approaches rely on single large-context LLM prompts, which frequently produce inconsistent, incomplete, or non-executable outputs due to the complexity of the generation task. This paper presents AppCompiler, a compiler-inspired, four-stage pipeline that decomposes application schema generation into sequential stages: intent extraction, architecture planning, schema generation, and validation with targeted repair. Each stage operates with a focused prompt and constrained output format, enabling the system to isolate and repair failures at the layer that produced them rather than regenerating the entire output. A cross-layer validation engine checks referential consistency across generated schemas and automatically resolves common classes of inconsistency. Evaluation across 20 test cases—10 representative product prompts and 10 edge cases including vague, conflicting, and underspecified inputs—demonstrates an 85–90% end-to-end success rate with a targeted repair rate of 30–40%, significantly outperforming single-prompt generation baselines in schema completeness and cross-layer consistency.