Hi, I'm Murali 👋

Implemented floor vs. non-floor pixel/region classification on indoor images using the CMM dataset and Support Vector Machines. Compared three feature-engineering approaches: RGB-only pixels, RGB plus spatial coordinates, and KMeans-based region-level features. Method 3 (KMeans regions) achieved ~92.5% test accuracy with the fastest training.

Stage 1 (KAT-PatchTST, first author, accepted at IEEE APPEEC 2026): physics-informed BA-aggregate load forecaster combining Channel-Independent PatchTST and TimeXer cross-attention with a Kirchhoff conservation regularizer and ReLoBRaLo dynamic loss balancing. Attains 3.55% demand MAPE on the six-BA EIA-930 protocol using a 168-hour context (30% shorter than the published 240-hour baseline) at ~0.6M parameters (3-10x leaner than comparators). Stage 2 (forecast-conditioned operational feasibility analysis on the BA-interchange network) and Stage 3 (learning a conditional grid-partition policy that minimizes allocation failures) are in development, targeting ICLR.

Implemented a distributed system using MPI (Message Passing Interface) in C++ to parallelize log file parsing and anomaly detection across multiple nodes, reducing analysis time by 70% for large-scale server logs. Integrated parallel reduction techniques for aggregating metrics like error rates and response times, enabling real-time monitoring and scalable debugging in cloud environments.

Developed a Monte Carlo simulation to estimate stock portfolio values, modeling returns with Cholesky decomposition. Simulated 100 portfolio projections over 100 days to assess risk and future performance.

Developed a neural option pricing system integrating deep sequence architectures with GARCH volatility, benchmarked against Black-Scholes, enabling quantitative comparison between ML based and Analytical pricing. Built an Interactive Dashboard with heatmaps, sensitivity analysis (price vs spot, volatility) and Multiple Models Trains with different features to visualize pricing behavior and model errors across market conditions.

Built an AI-powered Flask app that generates audio descriptions from images using Salesforce's BLIP image captioning model, combined with SVM-based floor classification (RGB-only, RGB+spatial, and KMeans-region feature strategies on a self-built dataset) to enhance environmental awareness. Converts uploaded images to speech using a text-to-speech engine, making it a practical assistive tool for visually impaired users.