Dawini

Inspiration

Dawini started as our graduation project—an AI-powered Arabic mobile healthcare assistant aimed at improving health accessibility and literacy in the Arab world. Initially, our focus was on providing personalized medical guidance for Arabic speakers, using AI to help users understand symptoms and navigate healthcare in our native language.

However, as the war in Sudan intensified, displacing millions and shutting down hospitals, our mission evolved. We recognized that Dawini could be more than just a technical achievement; it could serve as a vital lifeline. With healthcare infrastructure collapsing and misinformation spreading, Sudanese people needed a tool that comprehended Arabic and offered accurate, localized medical support. Dawini became a mobile assistant that helps users describe symptoms, receive relevant guidance, identify medical specialties, and find nearby doctors, ensuring accessible and trustworthy support for displaced families and those in remote areas.

What it does

Dawini is an Arabic-language AI health assistant that allows users to:

• Describe symptoms naturally in Arabic using a chatbot interface.
• Provides AI-generated, context-aware medical insights based on symptom description.
• Predicts the appropriate medical specialty through a structured health form.
• Recommends nearby doctors using GPS data and a structured doctor database.
• Automatically generates medical reports in PDF format, securely stored and shareable.

How we built it

• Frontend: Developed using Flutter for a smooth, cross-platform mobile experience.
• Backend: Built with Flask and FastAPI, integrated with Firebase for authentication and Firestore storage.
• AI Models: Aya Expanse 8B for Arabic symptom analysis, fine-tuned with domain-specific data using LoRA.
  CNN-BiLSTM classifier for medical specialty prediction using AltibbiVec embeddings.
• Report Handling: Cloudinary used for secure PDF storage and sharing.
• Deployment: AI models deployed via Lightning Studio with token-protected API access.

Challenges we ran into

• Scarcity of high-quality Arabic medical datasets for model training.
• Computational limitations when fine-tuning and deploying large LLMs like Aya Expanse 8B.
• Handling Arabic language complexity, dialects, and informal symptom descriptions.
• Doctor database built with simulated data, limiting real-world validation.
• Addressing data privacy and ensuring secure storage and communication within app constraints.

Accomplishments that we're proud of

• Built a culturally relevant Arabic AI health assistant that delivers symptom analysis and personalized doctor referrals.
• Fine-tuned a large Arabic LLM (Aya Expanse 8B) for accurate and safe medical symptom analysis.
• Achieved 77.35% accuracy and a 0.55 semantic match score in specialty classification using a hybrid CNN-BiLSTM model.
• Integrated both AI models into a fully functional, dual-role healthcare mobile app.
• Designed and implemented an intuitive Flutter interface for both patients and doctors.
• Prepared a high-quality Arabic medical dataset through extensive preprocessing and category restructuring.

What we learned

• Gained experience in fine-tuning Arabic LLMs for sensitive domains like healthcare.
• Understood the challenges of modeling overlapping medical specialties and user variability in input.
• Managed project constraints involving limited resources and real-world health application needs.
• Developed strategies for dataset preparation, prompt engineering, and response refinement to improve safety and clarity.

What's next for Dawini - داويني

• Build a real-time, location-based database of doctors, hospitals, clinics, dialysis centers, and pharmacies, and collaborate with Sudanese healthcare providers to keep it updated and accurate.
• Implement severity scoring and first-aid suggestions for urgent cases.
• Add a volunteer doctor chat feature, where verified doctors can offer remote consultations to users—especially useful during ongoing conflict and displacement.
• Integrate a feature that helps users find free or subsidized medications from local charities, NGOs, or public health centers when they cannot afford to buy them.

Built With

• cloudinary
• fastapi
• firebase-firestore
• flask
• flutter
• lightning
• python