Nov 2024 – Sep 2025 · 10 mos · New Delhi, Delhi, India · Remote

• Developed Anomalyze, a deep learning system for automated nuchal translucency (NT) measurement from first-trimester fetal ultrasound scans to support early detection of fetal anomalies.
• Built a prototype using LangChain + RAG to retrieve relevant prior ultrasound cases and assist clinicians with automated suggestions for risk assessment.
• Implemented U-Net, ResU-Net, and Attention U-Net for NT region segmentation and EfficientNet & Vision Transformers for risk scoring using TensorFlow and Keras.
• Designed preprocessing and augmentation pipelines (CLAHE, denoising, ROI extraction, GAN-based augmentation) in Python to improve robustness across diverse imaging conditions and limited datasets.
• Developed evaluation protocols with K-fold cross-validation, AUROC, IoU, and Dice Score for clinically reproducible results.
• Integrated PostgreSQL and MongoDB for structured and unstructured data storage; built vector database indexes to enable embedding-based search for similar cases.
• Applied NLP techniques to extract structured information from radiology reports and assist in automated report generation.
• Developed FastAPI endpoints for real-time model inference and deployment on hospital systems; tested models on embedded/edge devices, optimized via quantization and pruning for low-latency inference.