Building safer and scalable products for payment in India 🇮🇳

◦ Model Interpretability: Researched about Global Model-Agnostic Methods with rule based (if-else) outputs to explain model’s decision makings. Finally implemented the paper on GlocalX Algorithm (2021).◦ Scoped Rules Package: Created and deployed a python package having rule based model interpretability techniques like MAGIE and GlocalX at one place which is getting used across many data science teams in Rakuten

• Researched the various SOTA Model Compression Techniques and their implementation for resource limited devices• Tested the techniques like Quantization, Weight Pruning, Knowledge Distillation etc and their results in TF 2.0 and Pytorch • Implemented the paper “Dynamic Sparse Training, ICLR 2020” with a POC for Image Classification model in Pytorch

• Developed a Diabetic Retinopathy Classifier with 85,000 retinal images from Kaggle and Messidor dataset• Built a CNN architecture from scratch after iterative exploration, used transfer learning with few state-of-the-art CNN models• Improvised the performance by Stacking Ensemble of multiple CNN models and Kfold Cross-Validation, increasing accuracy by 7%

• Developed a pipeline of multiclass instance segmentation and classification model using Mask RCNN architecture in Tensorflow and Keras• Explored about various statistical models and integrated them to achieve a test accuracy of 90.7%• Created an automation image labeling tool using pyautogui library of python to reduce the manual labor