The care setting
Maternal-health clinicians in under-resourced hospitals lose valuable time to documentation, often while working across languages and inconsistent connectivity. The goal was to capture a spoken visit without forcing the clinician to become a data-entry operator.
- Designed for multilingual clinical speech and constrained workflows.
- Kept clinicians in control of the generated structured record.
- Prioritized usable output over free-form model responses.
Mobile voice capture to structured records
I built the React Native workflow and fine-tuned Whisper with LoRA on locally collected Urdu speech data. The resulting pipeline transcribes spoken clinical notes, maps them into a validated schema, and stores each encounter for review and follow-up.
- Fine-tuned Whisper with LoRA on locally collected Urdu speech data.
- Generated more than 500 structured records at 96% accuracy during a seven-month hospital deployment.
- Reduced clinician documentation time by 60%.
Risk detection grounded in guidance
A hybrid retrieval system checks patient answers against expert-curated clinical guidance instead of relying on unconstrained model memory. The service then surfaces red flags for clinician review.
- Used Elasticsearch retrieval over curated medical content.
- Raised high-risk condition detection from 7% to 40%.
- Surfaced more than 300 risk alerts.
From prototype to deployment
The mobile and backend pipelines were deployed as autoscaling microservices on GCP Cloud Run, with automated delivery designed for unpredictable clinic traffic.
- Shipped zero-downtime blue-green releases through GitHub Actions.
- Provisioned reproducible infrastructure as code with Pulumi.
- Published the system evaluation at ACM UbiComp 2026.