ONESTRUCTION, Inc., a construction technology startup, with technical advisory from the AWS Generative AI Innovation Center (GenAIIC), built Ishigaki-IDS, a foundation model specialized for construction industry BIM workflows, as part of GENIAC Phase 3. BIM is a digital representation of a building’s physical and functional characteristics. Japan’s construction sector faces a persistent labor shortage, and BIM is promoted nationally, but adoption requires specialist knowledge. Ishigaki-IDS lowers the barrier for non-specialists to review and manage attribute information in IDS, an XML-based standard defining information attached to and validated against a BIM model. The company faced three challenges: data scarcity, as IDS was published in 2024 and construction has limited public web content; the lack of a comparable public dataset to other domains trained on billions to hundreds of billions of tokens; and shallow collected web data. The post describes a reusable pattern for building specialized AI models where data is scarce, covering synthetic data generation, a three-stage training pipeline (CPT, SFT, RLVR), verifiable rewards for structured output generation, and distributed training on Amazon EC2 P5en instances with AWS ParallelCluster.