Using system-one models inside high-throughput data pipelines