• 

  SanRaksha

• 

  Home Screen

• 

  Graph in User Checkup Data

• 

  Risk Prediction

About the Project – SanRaksha: Safeguarding MOMs

Inspiration

In rural India, over 60% of pregnant women are anaemic, and many high-risk pregnancies go undetected due to lack of proper monitoring. ASHA workers, who form the backbone of maternal healthcare in villages, often operate with limited tools and rely heavily on manual records and experience. We were inspired to build a system that enhances their capabilities by combining local wisdom with intelligent, data-driven insights. SanRaksha aims to detect and prevent complications early, especially in areas with limited connectivity and medical infrastructure.

What We Built

SanRaksha our Misson for Optimal Motherhood is a full-stack maternal health monitoring system. It includes an Android app designed for ASHA workers that works offline and uses an on-device TFLite model for real-time risk prediction. The app provides personalized dietary and clinical advice in multiple languages. Complementing the app is a PHC (Primary Health Center) web portal for doctors, allowing them to view risk alerts, sync field data, and trigger model retraining when needed.

Data & Model Challenges

Working with maternal health data presented several challenges. Many of the datasets we sourced were incomplete, noisy, and clinically imbalanced. We had to carefully handle missing values using domain-aware imputation strategies, ensuring no data leakage during preprocessing. Defining what constitutes a "high-risk" pregnancy was not straightforward—clinical thresholds for indicators like hemoglobin or blood pressure vary regionally, requiring cross-referencing with medical guidelines. We also faced the challenge of balancing model performance with interpretability, especially since the end-users are health workers without a technical background. Therefore, we emphasized transparency and built simple explanations into the predictions.

Integration Hurdles

Bringing machine learning into real-world, low-resource settings is a challenge in itself. Deploying models on low-end Android phones meant working within tight memory and performance constraints. We optimized our model using TensorFlow Lite and ensured it could run completely offline. Handling updates in a learning setup required careful design—models had to be synced efficiently whenever internet access was available. Building an infrastructure that is functional in remote areas was a technically demanding but essential part of this project.

What We Learned

This project taught us how to design AI systems that are not just accurate, but also practical and trustworthy in sensitive, real-world applications. We learned the importance of domain alignment when working with healthcare data, and how easy it is to unintentionally introduce bias or leakage if standard ML workflows are not adjusted. We also gained experience in deploying local models and integrating federated learning into mobile apps—an area that's still evolving and filled with real engineering trade-offs.

What’s Next

We are currently in discussion with local healthcare NGOs to pilot SanRaksha in actual field conditions. Our next steps include improving regional dietary recommendations, integrating SMS-based alerts, and expanding support for more languages. We also aim to build a federated learning infrastructure to better to handle asynchronous updates and data versioning in the field, along with privacy preservation

Built With

• fastapi
• kotlin
• litert
• python
• scikit-learn
• tensorflow
• verstack
• xgboost