Research Assistant · RMSoEE, IIT Kharagpur
Prototype
?
Draw a batch, then scan it.
Stylized re-creation of the real fraud-detection model — a 5-class CNN trained from scratch on a proprietary dataset of 500 high-resolution saffron strand images, reaching 71% validation accuracy under severe low-data constraints. The scanner above makes mistakes at exactly that rate.
?
Stylized benchmark mirroring the real result — Random Forest set the pace for localized saffron price forecasting, while XGBoost topped agricultural yield optimization. All four models were hyperparameter-tuned over the same historical records.
Highlights
- Multi-Objective Agritech Framework: Led a data-driven agritech research project targeting three operational milestones: saffron market price prediction, crop yield optimization, and adulteration/fraud detection.
- Tabular Modeling & Benchmarking: Configured, hyperparameter-tuned, and cross-benchmarked four predictive architectures—Multiple Linear Regression, Support Vector Machines (SVM), XGBoost, and Random Forest—over historical records; Random Forest delivered the best localized price forecasts, while XGBoost topped yield-optimization accuracy.
- Computer Vision for Fraud Detection: Designed and coded a 5-class deep Convolutional Neural Network (CNN) image classifier from scratch to automatically detect, screen, and flag saffron adulteration and fraudulent crop mixtures.
- Proprietary Dataset Formulation: Built, cleansed, and curated a proprietary crop-monitoring image dataset of 500 high-resolution saffron strand images to support model training under strict low-data constraints.
- Model Training & Generalization: Mitigated overfitting under tight data limits to reach 70% training and 71% validation accuracy, validating robust feature extraction on a sparse image dataset.
- Agritech Decision Support & Impact: Distilled explainable machine learning insights into an agricultural decision-support framework, giving supply-chain operators data-driven levers to protect product authenticity and improve cost efficiency in saffron cultivation.
Technologies
- Random Forest
- XGBoost
- SVM
- CNN
- Computer Vision
- Python