We built Synthpylon because privacy-blocked data was killing our models
Based in Columbus, OH. Founded 2025.
Before Synthpylon, Hannah Mueller spent three years on the ML infrastructure team at a mid-size healthcare analytics firm in Columbus, OH. The job was model iteration: retrain, evaluate, deploy. Every improvement proposal ran into the same wall. The interesting data, patient outcomes, rare diagnoses, treatment sequences, sat behind HIPAA access controls that the contracted ML engineers couldn't cross.
The workarounds accumulated. Anonymized exports broke downstream pipelines because k-anonymity shifts distributions. Off-the-shelf synthetic tools either required a statistics PhD to configure or produced outputs that failed the team's own fidelity checks. One vendor's output had negative lab values. Another silently dropped foreign key relationships between patients and encounters.
Hannah started a generation pipeline as a personal side project in late 2024. It connected directly to database schemas, fit a Gaussian copula over column marginals and pairwise correlations, and produced outputs that actually passed the team's quality gates. The first version was a single Jupyter notebook and ran on her laptop.
Co-founder Daniel Osei, then finishing a PhD in applied statistics at Ohio State, joined in early 2025 after Hannah shared the evaluation methodology with him. They formalized the statistical approach, added the differential privacy budget controls, and incorporated Synthpylon in July 2025. They had not planned to start a company. The alternative was watching the same pipeline-blocking problem repeat across every team they talked to.
The team
Hannah Mueller
CEO and Co-Founder
Three years on the ML infrastructure team at a healthcare analytics firm before founding Synthpylon. Columbus native. Ohio State CS, 2020. Spent most of her career watching useful models fail to ship because the data needed to train them couldn't leave the building.
Daniel Osei
CTO and Co-Founder
Applied statistics PhD candidate at Ohio State, left ABD to co-found Synthpylon. His research focused on copula-based dependency modeling for tabular data. Before the PhD, two years on the risk modeling team at a Columbus fintech. Owns the statistical core and the fidelity scoring methodology.
How we think about this
Privacy is a constraint, not a feature
We don't sell "privacy-first AI." We built a tool where privacy is structural. Not touching real records isn't a setting you turn on, it's how the system works.
Fidelity scores are non-negotiable
Synthetic data that doesn't actually help your model is worse than nothing. It wastes your time and builds false confidence. We ship a fidelity report with every dataset.
Tabular data only, for now
We do not synthesize free-text, images, or audio. We do tabular structured data for ML training and analytics. That is a narrow lane and we are going deep on it. Time-series is on the roadmap. Unstructured data is not.