Healthcare data is often fragmented across hospitals, clinics, reports, images, and mobile applications. Swasth AI was designed as a research-driven attempt to bring those signals together and produce more complete diagnostic support.
The model-fusion concept assigns different data types to models suited to them: convolutional neural networks for medical images, sequence models for time-series vitals, and transformers for clinical notes. Their outputs are then combined into structured health insights rather than treated as isolated predictions.
Engineering priorities
- Keep each model's contribution visible and reviewable.
- Normalize inconsistent input data before inference.
- Present outputs as decision support, not autonomous medical advice.
- Build mobile-first flows that make complex information understandable.
I led frontend delivery for the product and contributed to the wider research and system design. The work was published as "Swasth AI: Unifying India's Fragmented Healthcare System Through AI-Powered Diagnostics and Model Fusion" in the proceedings of IEEE ACROSET 2025.
The project reinforced an important principle for me: high-stakes AI systems need strong product design, traceable reasoning, and explicit limits just as much as they need capable models.