Varun Kadekar

Auto Speech Recognition (ASR) System:

• Built post processing application for ASR system, using deep learning techniques
• Processed the raw text output from ASR system and punctuated it using retrained Bert based model
• Used customized regex to clean up data and NER based true-casing system to true-case the words.
• Used various techniques to convert text to digits and thus improved system's accuracy.
• Tech Stack used: Azure blob, deep learning using pytorch, RabbitMQ, Dockerized the module before productionizing

Financial Forecasting:

• One of the business pain areas is allocating the expense budget at the lowest level of every Cost Center or forecasting the margin for every Product. This task was automated using Machine Learning
• Techniques Used – Linear Regression with dynamic variable choice, spline data interpolation, univariate outlier detection, Ljung box test with box cox transformation, Gradient Boost, Neural Network and Arimax models
• Multi core processing ML scripts were deployed on High Performance Computing server

Anomaly detection - log data analysis:

• Use case was to identify the event distribution patterns in a dataset that do not conform to expected patterns and then identify if the non-conformance has led to a Service Request creation. Model was trained on 1.4 billion rows.
• Techniques Used: Chi-Sq, Random Forest, SMOTING, kProto Clustering, Auto Encoders

Customer Churn Analysis in Telecom:

• Churn Analysis of High-Speed Internet Customer. Customer Data was tracked for 9 months, validation done using 3 months data
• Algorithms Used: Random Forest and PCA , Logistic Regression

Adhoc Tasks:

• Cross Industry recommendation system with data from multiple domains along with the customer demographic details
• Worked on Information Quality workbench build to help with information insights with minimum human intervention, using open source framework like R, Python and R-Shiny.
• Data driven spelling correction without references to external dictionaries
• Multi variate Outlier detection using Regression, Auto encoders and clustering
• Built Geo based visualization by integrating Tableau and R for an optimization use case.

• Leading and delivering critical projects in the Telecom billing domain
• Subject Matter Expert for telecom billing
• Optimized operations leading to a savings of $35,000
• Built POC for anomaly detection based on call detail records, sentiment analysis of various product reviews