I build systems that read signal out of messy data.

Automated training pipeline that detects failures and recovers without manual intervention — retraining, rollback, and alerting built in.

Models don't fail loudly — they rot. Inputs drift away from the training distribution and accuracy slips for weeks before anyone notices. I built this so the pipeline watches its own data, catches drift statistically, and retrains itself before degradation ever reaches users — making reliability a property of the system instead of a fire drill.

A 1D CNN that detects apnea events from single-lead physiological signals, validated with leave-one-patient-out cross-validation so results hold on unseen patients.

Predicts pre-eclampsia risk from routine clinical features — one of the leading causes of maternal mortality worldwide. Built to flag high-risk pregnancies early enough for intervention to change the outcome.

Pre-eclampsia is one of the leading causes of maternal death, and the hard part is that it's often catchable — the signal sits in routine checkup data well before it becomes an emergency. I built this to surface that risk early, from features clinicians already collect, so intervention can happen while it still changes the outcome.