Projects & Blogs

Documenting my iterative process on creating a multi-model

Sharing my journey and experience on self-hosting and building this portfolio site using Eleventy, Bootstrap and Tailscale. While looking into the services that Tailscale offered, I stumbled upon Tailscale Funnel, which routes local hosted services, to the internet, enabling public internet access to it. Hence, I created a virtual machine on Proxmox with minimal resources, dedicated to serving this site.

Sharing my journey in setting up Proxmox Virtual Environment as a Hypervisor OS for virtualization needs, TrueNAS as a Network Attached Storage (NAS) OS, and Tailscale for secure remote access to these local services on the Proxmox Virtual Environment host.

A writeup on what I have done and learned in Ernst & Young as a Technology Risk & Compliance Auditor in Financial Services.

A comparative study of neural network models (LSTM, CNN-LSTM) for S&P 500 index prediction. The analysis evaluates multiple architectures against baseline strategies like buy-and-hold and naive forecasting, finding that while specialized models like LSTM with Batch Normalization and Regularization improve technical error metrics (MSE, MAE). However, when tested with a simulated trading, beating market returns remains a significant challenge requiring further architectural and strategy refinement.

This project involves building a credit score classification model using PySpark's machine learning library. By leveraging distributed computing, we compare the performance of Multilayer Perceptron, Decision Tree Classifier, and Random Forest Classifier to predict creditworthiness based on selected customer and financial features.

This project explores various capabilities of distributed computing across three distinct analytical domains by processing large datasets to perform descriptive statistics and clustering as well as image classification. The implementation includes Hadoop MapReduce jobs for weather data analysis on a dataset of hourly weather observations and unsupervised learning using Apache Mahout with several distance metrics, on a dataset of french plays. This work also showcases a scalable cat and dog classifier using the CLIP model within a Hadoop Streaming framework.

A comprehensive comparative study of time series forecasting methods across six diverse datasets (Gold, S&P 500, FTSE, USDC/USDT, StackOverflow questions, and cinema tickets). This analysis evaluates 10+ forecasting techniques—from naive baselines to ARIMA/SARIMA models—with quantitative performance metrics (MSE, RMSE, MAE), explores stationarity testing, seasonality decomposition, Monte Carlo simulation, and implements a simulated trading strategy.

Developing a Decision Support System to predict financial literacy and behaviour along with well being based on socioeconomic factors using a Bayesian network.