Synthetic data for healthcare AI: reverse inference, let clinicians own the pipeline

Anterior faces a common vertical AI constraint: they can't retain PHI. Their solution reverses the inference workflow. Starting from policy-defined labels and deterministic reasoning traces, they sample diverse scenarios from symbolic decision trees. This structured prior avoids LLM mode collapse. Then a coarse-to-fine LLM pipeline builds patient journeys, encounters, and

documents layer by layer, using clinician-guided skills. The key insight is that domain experts own the pipeline, not engineers. By modeling the pipeline as skills on an internal agent harness, clinicians add or modify document types without engineering changes. The data is generated in text, not PDF, relying on modern

parsers. A round-trip check ensures labels are correct by construction, eliminating expensive ground truth labeling. Results are promising: 90% of evaluation datasets are synthetic, and blind reviews show clinicians distinguish synthetic from real only 60% of the time. The real value is just-in-time data creation for new customer deployments and

edge-case simulation before production. Teams can model failure cases proactively rather than reactively. The tradeoff is fidelity vs. diversity. Synthetic data generation is a pipeline engineering problem, not a prompt engineering problem. The strongest signal is that AI adoption in regulated domains requires rethinking data ownership, not better models.