Projects
Sentiment Analysis on SBI YONO Mobile App Customer feedback
Analyzed over 50,000 recent customer reviews of SBI mobile app called “YONO” through sentiment analysis.
Identified positive, negative, and neutral remarks to better understand product’s quality and customer acceptance in the market. Clustered and extracted frequently used keywords from the comments and intimated corresponding team for improvements.
Model - Huggingface pretrained BERT model on English language
Model optimisation and hyperparamter tuning with help of Weights & Biases (wand)
Evaluation Metrics : F1-Score, Confusion Matrix
Machine Learning model to detect and forecast debit card fraudulent transactions
Built an ML model to evaluate the veracity of card transactions. With the aid of this model, the Bank's fraud monitoring team was able to keep the customer's account balance intact.
Extracting insights regarding debit card fraudulent transactions
Pre-processed 1 billion debit card transactions, categorised them as fraudulent or not, and extracted insights regarding the fraudulent transactions. With the help of these insights, Bank's fraud monitoring application monitor thousands of dubious transactions each day without affecting the balance in the customer's account.
Machine Learning model to predict early warning stress signals in p-segment loans
Created an ML model to learn about a customer's credit and repayment history and estimate the likelihood of nonpayment(NPA:Non-Performing Asset).
Algorithm : XGBoost
Baseline : SVM, LR
Evaluation Metrics : F1-Score, Confusion Matrix, Precision, Recall, AUC-ROC curve
Machine Learning model to identify eligible customers for pre-approved loans
Implemented an ML model for predicting Customer eligibility for pre-approved loan with no application fees, no documentation, no branch visit, and 24/7 availability.
Oracle Golden Gate data replication from one database to another
Filtered, replicated, and transformed data from one database to another using Oracle GoldenGate.
Android application to identify Customer Satisfaction on SBI branch visit
Designed and developed an android application that consumes Microsoft Azure pretrained ML models in order to recognise customer satisfaction with branch support based on facial expressions.
Rule-based debit card fraud detection and monitoring tool
Designed and programmed a Java application to detect suspicious transactions and send out alerts when transactions match predefined rules. Bank's fraud monitoring cell watches the alerts around-the-clock and takes necessary steps, such as verifying transaction legitimacy and barring the debit card in case of fraud, by consuming numerous APIs.
Impact: Application handles millions of transactions each day. Averted the expense of buying similar software off the market.
POC on Aadhar based account opening (finger print based authentication)
Designed and developed a Java application that helps SBI in touch brnach (fully digital, zero staff) to open accounts for new Customers. Authentication is done by capturing user finger prints and consuming webservice hosted by Unique Identification Authority of India (UIDAI). After succesful authentication, customer will be able to keyin the details relevant to account opening.
SMS-based tool to decrease SBI cash machine downtime
In order to serve 27,000 SBI branches in India and get 67,000 cash machines (ATM) operational, I designed and programmed an SMS-based solution in Java. The application processes close to 1 lakh SMS requests each day. System accepts SMS requests, checks centralised switch for cash machine's technical issues, and provides configuration file downloads to machines to get them up and running.
This application reduced average downtime of ATMs across India and halved the number of tele supporting staffs.
I'm happy of having created a system that enables people to use ATMs without experiencing any downtime, especially for those living in rural and mountainous areas of India, where mobile and internet banking are less common and ATMs are the only place to withdraw cash