Verticals: Collections, Fraud, UPI, & CX• Analyzed the P2P UPI transaction patterns to recommend periodic subscriptions to userwith a 30% Hitrate.• Reduced monthly cashback burn by 5 lakh using a rule-based predictor that analyzes userbehavior, identifying and converting 80% of abusers to the hard category.• Developed first-party fraud detection PoC for onboarding using tree-based models;classification power insufficient for production due to <0.01% anomaly rate.•… Show more Verticals: Collections, Fraud, UPI, & CX• Analyzed the P2P UPI transaction patterns to recommend periodic subscriptions to userwith a 30% Hitrate.• Reduced monthly cashback burn by 5 lakh using a rule-based predictor that analyzes userbehavior, identifying and converting 80% of abusers to the hard category.• Developed first-party fraud detection PoC for onboarding using tree-based models;classification power insufficient for production due to <0.01% anomaly rate.• Built a concurrent speech-to-text pipeline for B0 DPD customers with audiopreprocessing, speaker diarization, and quantized Whisper inference, generatinguser-level insights for collections.• Enhanced First Response Automation logic for specific tags, maintaining high precision incustomer queries using a combination of topic cluster modeling, GPT-matched questions,and keyword matching Show less

• Vectorised operations for fetching 2,50,000 Selfie Images, devised customgenerator functions to load Image data in batches and prevent RAM overflow• Fulfilled low latency of 350ms per image on Intel i5 CPU of complete pipelinefrom preprocessing to predictions• Used statistical analysis on the data labels, generated risk ranking bins basedon model’s prediction probabilities, accomplished 80% Probability buckets ofcustomers, which can default.• Designed Industrial… Show more • Vectorised operations for fetching 2,50,000 Selfie Images, devised customgenerator functions to load Image data in batches and prevent RAM overflow• Fulfilled low latency of 350ms per image on Intel i5 CPU of complete pipelinefrom preprocessing to predictions• Used statistical analysis on the data labels, generated risk ranking bins basedon model’s prediction probabilities, accomplished 80% Probability buckets ofcustomers, which can default.• Designed Industrial Image classification models, and Formulated complete Pipelines on a large consumer dataset. Show less

• Designed a Pipeline which involved improvement of data distribution,extraction of facial features and classification into the 7 basic emotions.• Improved class-wise accuracy using various performance metrics such asclassification accuracy, precision, recall, F1 score and AUC-ROC.• Solved the Model Bias due to Imbalance of Classes using Data resamplingtechniques, synthetic data generation using SMOTE on the FER-2013 datasetwhich contains 30,000+ grayscale images… Show more • Designed a Pipeline which involved improvement of data distribution,extraction of facial features and classification into the 7 basic emotions.• Improved class-wise accuracy using various performance metrics such asclassification accuracy, precision, recall, F1 score and AUC-ROC.• Solved the Model Bias due to Imbalance of Classes using Data resamplingtechniques, synthetic data generation using SMOTE on the FER-2013 datasetwhich contains 30,000+ grayscale images collected from Google image search.• Work was done under the supervision of Prof. Pramod Kumar Singh. Show less

• Trained a Computer Vision model to detect and recognize the number plates of bikes and delivered 93% accuracy on a 5000-image dataset.• Removed noise using grayscale conversion, dilation and erosion and Gaussian blur to smoothen out edges and binarized the images.• Detected yellow numberplates using OpenCV and improved the region of interest by 60% with the help of bitmasking of colour in HSV range.• Eased the pairing of IoT and vehicle to handle fleet management. Provided a… Show more • Trained a Computer Vision model to detect and recognize the number plates of bikes and delivered 93% accuracy on a 5000-image dataset.• Removed noise using grayscale conversion, dilation and erosion and Gaussian blur to smoothen out edges and binarized the images.• Detected yellow numberplates using OpenCV and improved the region of interest by 60% with the help of bitmasking of colour in HSV range.• Eased the pairing of IoT and vehicle to handle fleet management. Provided a secure and cost-effective solution using Open-Source libraries.• Extracted text using EAST text detector, performed OCR using Tesseract with LSTM neural network model. Show less

Department of Online Ventures And Freelancing of E Cell.