Medi-Shield: Offline AI-Based Emergency Medical Support

Inspiration

The inspiration for Medi-Shield came from thinking about how medical emergencies actually unfold in real life. In many cases, emergencies happen in places where internet connectivity is weak or unavailable, such as rural areas, highways, or during natural disasters.

In these situations, people often panic and do not know what immediate steps to take while waiting for medical help. Most existing medical or AI-based applications depend on cloud services, which makes them unreliable exactly when they are needed the most. This motivated us to explore the idea of an offline, on-device AI system that can provide basic emergency support without relying on the internet.

What it does

Medi-Shield is an offline AI-based emergency medical support application designed to assist users during the initial critical moments of a medical emergency.

The app allows users to: 1.Describe symptoms using voice input in their local language 2.Receive immediate, step-by-step guidance without internet access 3.Get basic first-aid and stabilization instructions 4.Automatically trigger an emergency call if the situation is critical

The system does not diagnose diseases or prescribe medicines. Its purpose is to support users until professional medical help becomes available.

How we built it

Medi-Shield is designed as a fully on-device system, without any dependency on cloud APIs. At a high level, the workflow is:

User Voice Input → Local Speech-to-Text → On-device Language Model → Offline Medical Protocols → Text and Voice Guidance

1.Whisper (local) is used for speech-to-text to support multiple languages. 2.A quantized small language model (DeepSeek-R1-Distill) is used for reasoning and decision support. 3.The RunAnywhere SDK is used to manage efficient model execution on mobile hardware. 4.All medical guidance is based on preloaded, doctor-validated protocols.

All processing happens locally on the device to ensure privacy, low latency, and offline reliability.

Challenges we ran into

One of the main challenges was working within the hardware constraints of mobile devices. Memory and compute limitations required careful model selection and quantization.

Another challenge was defining clear ethical boundaries. We intentionally limited the system so that it does not attempt diagnosis or medication recommendations.

Designing for offline usage also meant that all medical information had to be stored locally, which required careful consideration of content scope and size.

Accomplishments that we're proud of

1.Designing a complete on-device AI pipeline without cloud dependency 2.Creating a system that prioritizes privacy and reliability in emergencies 3.Clearly defining the scope to avoid unsafe medical claims 4.Demonstrating how small language models can be used effectively on mobile devices We are particularly proud of keeping the solution practical and realistic.

What we learned

Through this project, we learned that building on-device AI requires a very different approach compared to cloud-based systems.

Key learnings include:

1.The importance of efficiency and model optimization 2.How system design changes when internet access is removed 3.Why scope control is essential in healthcare-related applications 4.How data engineering and system constraints shape real-world AI solutions

This project helped strengthen our understanding of practical, constraint-driven AI design.

What's next for Medi-Shield: Offline AI-Based Emergency Medical Support

Future improvements for Medi-Shield include:

1.Expanding support for more regional languages 2.Improving the offline medical knowledge base 3.Integrating with wearable devices for basic vitals monitoring 4.Adapting the system for disaster response and rural healthcare programs

The long-term goal is to continue exploring how offline AI can make critical systems more reliable and accessible.

Built With

• kaggle
• local-whisper-speech-to-text
• python
• quantized-deepseek-r1-distill
• runanywhere-sdk