Abhishek
Tiwari

Automated plant disease diagnosis using InceptionNetV3 and EfficientNetB3 on 87,000+ leaf images across 38 classes. Enhanced with histogram equalization, Gaussian blur, edge detection, and real-time augmentation. EfficientNetB3 converged in just 5 epochs; InceptionNetV3 achieved top test accuracy over 12 epochs — proposing a scalable precision agriculture solution.

Investigated Random Forest, SVM, LSTM, Bidirectional LSTM, and a novel Hybrid RF+LSTM model for time-series groundwater level forecasting from monitoring station data. Preprocessed with interpolation, normalization, and lag features. The Hybrid model achieved the best trade-off between structured and sequential learning, outperforming standalone approaches.

Transformer-based multi-task framework using TransUNet for automated PE segmentation and classification on CTPA scans from the FUMPE dataset (35 patients, 200+ DICOM slices). Outperformed SCUNet++ (Dice 83.4%) and Swin-UNet via global attention and cross-attention refinement for capturing small, scattered emboli with minimal false positives.

Custom NLP Transformer built with a proprietary Emotion-Embedding and Modified Attention layer. Optimized stability via dropout tuning and label smoothing.

An automated service that monitors ML model performance and triggers retraining when accuracy drops below a threshold. Uses Celery Beat for periodic health checks every 30s, RabbitMQ as message broker, and FastAPI for REST API endpoints.