Clinical AI Assistant
Self-hosted LLaMA and on-device speech, tested across Indiana hospitals under HIPAA constraints.
- Software Engineer, AI Research
- Purdue University
- 2024 to 2025
- Research
Problem
Clinical AI has to be useful without being careless. Privacy, latency, and interpretability matter as much as the model does, and a nurse mid-shift will not wait on a round trip to someone else's API.
Approach
I built a clinical assistant with hard privacy constraints: nothing leaves the device that doesn't have to. Local speech-to-text, self-hosted models, and an interface a nurse could use one-handed while doing something else.
What I built
A Flutter and TypeScript clinical assistant tested across Indiana hospitals, running self-hosted LLaMA 3.2 and an on-device, HIPAA-compliant speech-to-text pipeline.
Architecture
Flutter and TypeScript application layer built around voice-first clinical workflows: tasks, reminders, and patient vitals.
Self-hosted LLaMA 3.2 with a zero-API, on-device speech pipeline, so protected health information never leaves the device.
Deployed and tested across Indiana hospitals, then presented at the Purdue Spring Research Conference.
Impact
Cut nurses' documentation overhead by roughly 40% in hospital testing, and was presented at the Purdue Spring Research Conference.
~40%
