Harshad Saglani

Actuarial departments in insurance companies determine the rates customers pay and receive for their policies. Actuaries review summary reports of premiums, losses, cancellations, and conversions.

The data for these reports came from multiple systems, and summary reports didn't match exactly with the operational reports from those systems. The discrepancies were caused by differences such as using effective dates versus calendar dates, back-dated transactions, and system bugs.

We needed all the reports to match to be confident that our reporting and operational systems were in good health. To accomplish this, I wrote many programs in Oracle PL/SQL that scanned the transactions from systems and compared them with the reporting data and vice versa.

This utility was a huge success and saved at least one week every month for 1-2 people who used to manually verify the discrepancies in the reports. This report also led to a high level of confidence in the system, continually uncovered system issues, and was very helpful in testing the overall health of the systems.

In this project, I tested the bulk processing and automating data pipelines with Amazon DynamoDB, AWS Lambda, Amazon S3, Boto3, and PostgreSQL. I also tested various approaches for data transformations using SQL, pandas, and Spark (partitioning using parquet files) and compared the performances of different approaches.

A handwritten digest classifier based on a deep learning model trained on the Modified National Institute of Standards and Technology (MNIST). This application uses the fastai library, which uses PyTorch internally. The user can draw any digit in this application, and the application identifies it.

This project has a neural network (deep learning) model and a UI where users can draw a digit and see the result (prediction). I worked on this project to gain a working knowledge of developing and deploying a machine learning model and using it from a UI, data pipeline, or workflow.