Rebase

Every company eventually outgrows its data infrastructure, but schema migrations remain largely manual. While frontier LLMs achieve high field-mapping accuracy, they consistently fail on structural transformations like merges, splits, nested extraction, and relationship preservation—the hardest and most critical part of migration.

Rebar is an RL training stack built to solve that gap. We generate schema migration tasks, execute model-generated transformations, and grade results across field accuracy, relationships, and structural correctness. Those signals feed back into online reinforcement learning, continuously improving migration performance.

Our goal is fully autonomous data migration: models that can understand two schemas, generate and execute the required transformations, and return verified results without hand-written mappings or engineering intervention.