project file / apr 2026
MedKit
An offline two-device experiment for field triage, pairing an Android phone with an on-device model on a Snapdragon laptop.
- my role
- Builder, with a hackathon team
- status
- Research prototype
- field
- Hackathon prototype / edge AI
- tools
- Python · Gemma · LoRA · CLIP · ChromaDB
01 / the problem
Why this exists
Field teams may not have reliable internet access. MedKit tested whether a phone could collect an image and short description and pass them over local Wi-Fi for offline model assistance.
02 / the build
What I worked on
Prototyped phone-to-laptop transfer and a local processing flow for injury images and patient descriptions.
Built image and text retrieval with CLIP and ChromaDB and evaluated LoRA tuning of Gemma for concise guidance.
Documented integration and model-quality limits found during hackathon testing.
03 / engineering notes
Choices that mattered
- Keep the prototype offline and make the data path explicit rather than depending on a remote model.
- Treat output as an experiment to evaluate, not as a clinical decision or a validated medical tool.
04 / reflection
What I took away
The prototype showed promise as an offline workflow and also exposed how much validation would be required before guidance like this could be trusted in practice.