Medical AI Researcher | Keras Team Contributor (Google) | Ex-DICOM AIML Intern @ UniDoc Healthcare
Passionate about trustworthy, multimodal AI for healthcare. Building production-grade systems for 3D segmentation, adversarial robustness, federated learning, and assistant tools.
🔭 Current Focus: Integrating everything into MedAI Assistant – a voice-enabled, robust multimodal system (segmentation + report gen + RAG + foundation models).
🌱 Latest Win: Merged RandomElasticDeformation3D layer to Keras-Hub (PR #2419) – boosting 3D augmentation for Gemma/TPU workflows.
| Project | Description | Stars | Key Results |
|---|---|---|---|
 Medical-Segmentation-Decathlon |
TPU-native SOTA framework for MSD (10 tasks). UNET++/TransUNet/SwinTransUNet + 3-stage OHEM. | |
73-94% Dice across tasks • Sliding-window inference |
| Adversarial-Defense-MedAI (WIP) | Robustness for med seg: FGSM/PGD/CW attacks + defenses (adv training, StyleGAN aug). Config-driven. | – | Defended Dice recovery demos coming |
| Project Swasthya | Privacy-first federated enviro-health system (Flower FL + AWS MLOps). Unified ERD/UML blueprints. | – | Phased MVP → Causal Inference |
| MedAI Assistant (In Progress) | Multimodal med assistant: Seg + Report Gen + RAG + Voice. Culmination of all work. | – | Jan-Jun 2026 MVP |
| Core Frameworks | Medical & MLOps | Generative/Robustness |
|---|---|---|
| TensorFlow • Keras • PyTorch | NiBabel • PyVista • TPU • AWS/GCS • Flower FL | StyleGAN • Adversarial Training • DP |
- LinkedIn: ashpak-shaikh-88a7372b0
- Email: [ashpakshaikh26732@gmail.com]
Contributions welcome! 🚀