Federated learning
Clinical data is siloed by institution, jurisdiction and law. Pooling it is slow and often impossible, and a model trained at one hospital tends to learn that hospital’s scanners and coding practices rather than the thing you wanted it to learn.
At Mayo Clinic Platform I build the platform that moves the model instead of the data:
- The model is sent to each institution.
- Each institution trains it on its own records. Only parameters leave.
- An aggregator combines what came back.
- Repeat.
The platform covers secure data environments with access to 13.6 million patient records and 3.9 billion medical images. I also direct the replacement of the legacy Mayo Clinic Cohort Visualizer, which identifies cohorts across sites without centralising the records. As technical lead in NVIDIA’s NVFlare clinical engineering stakeholder group, I translate clinical and regulatory requirements into platform requirements.
The hard part is not the averaging. It is identity, consent, audit, and making the system fast enough that a clinician will use it while a patient is in the room.
Before Mayo
- Tatari, Senior Machine Learning Engineer (2018–2025): multi-armed bandit systems for media optimisation, and mentoring engineers into senior roles.
- TrueCar, Software Engineer 3 (2015–2018): search, APIs, and the Python platforms behind them.
- Florida Department of Economic Opportunity, Architect and Engineering Lead (2014–2015).
Tools
Python, PyTorch, NVIDIA NVFlare, Kubernetes, Terraform, Airflow, Spark and Databricks, PostgreSQL.