♦ Developed UI to provide risk managers the ability to visualize time series of credit risk exposures to better understand the efficacy of risk model with respect to its VaR metrics due to unexpected behavior after COVID. Technologies: Python, Streamlit, Pandas

♦ Built a reporting tool on Redshift that allowed the Risk managers to easily review log files from the model for a specific portfolio to confirm there were no errors or warnings. Technologies: Redshift, Python, Streamlit

♦ Cost-optimized use of Redshift by maintaining CDS Index net exposures and positions in S3 for UI queries. Technologies: Redshift, S3, boto3

♦ Developed framework to dynamically detect and visualize missing data and significant spikes in time series used in firm’s risk modeling workflow, and generated summary statistics for each series. Technologies: Python, Pandas, NumPy

♦ Developed new AI application that uses reinforcement learning to easily upload and classify new datasets interactively ranging in size from 10 GB – 100 GB and provided initial results to user in under one minute. Technologies: Python, Keras, Pandas

♦ Built a Python framework to allow engineers to leverage Gym to quickly validate the outcome of their neural network model by running 1000 iterations of simulations in a test environment. Technologies: Python, Gym, Keras.

♦ Designed a Jupyter notebook to provide transparency to each stage of the simulations executed in Gym that can plug and play with any of the company’s deep learning models, including Keras, PyTorch.

♦ Developed RESTful Python microservice for a financial analysis dashboard that displayed market data and volatility for a Ticker. Technologies: Python, Docker, Vagrant CentOS VM, and Ansible

♦ Cut reconciliation time to 5 minutes from 1 day by replacing a Java application that used a line-by-line reconciliation that did not scale with a Python script leveraging Pandas. Technologies: Python, Pandas, cx_Oracle, xlsxreader

♦ Benchmarked 8 machine learning algorithms to determine which one performed the best on a representative data set from the firm’s production portfolio analysis suite which the firm used to select the best algorithm for their core product. Outcome: SVM most accurate but didn’t scale due to number of features required. Naïve Bayes performs well, CPU/Memory intensive due to text-to-numeric conversions. Neural Net requires a lot of configuration and might be more complex than needed. Technologies: Decision Tree, Naïve Bayes, SVM, Random Forest, Gradient-Boosted Tree, K-Means Analysis, Neural Network, KNIME