Embedded Files
Current Projects
Neuro Symbolic AI
Researching Neuro-Symbolic AI, combining symbolic reasoning, causal analysis, counterfactual simulation, semantic graph rewriting, DreamCoder-based program synthesis, and MAML for GNNs to enable adaptive, multi-domain intelligent agents.
Completed Projects
InstaBot AI
An automated bot that generates prompts and captions using DeepSeek-R1, creates AI images with Flux, and auto-posts them on Instagram, streamlining content creation.
github: InstaBot AI
Chatbot and Feedback Tool
Developed a website-integrated chatbot using Grog API with multi-turn conversation support and Phoenix-based observability. Built a Streamlit feedback tool for annotations and an analytics dashboard for trace export and reporting.
github: Chatbot
Simulating Shallow Water Equations (SWE) Using Physics-Informed Neural Networks
I'm working on simulating the Shallow Water Equations using Physics-Informed Neural Networks (PINNs). The key challenge is adapting the geometry to a spherical domain with periodic boundary conditions to realistically represent Earth-like behavior.
Daily Prediction of Terrestrial Water Storage Anomalies Using Physics-Informed Neural Networks
This project focuses on predicting daily Terrestrial Water Storage Anomalies (TWSA) across the Amazon basin using Physics-Informed Neural Networks (PINNs). By combining satellite data from the GRACE (Gravity Recovery and Climate Experiment) mission with atmospheric reanalysis data from ERA5, the project incorporates physical principles into machine learning models to improve the accuracy of water storage predictions. This approach aims to provide a better understanding of regional water storage dynamics and support more effective water resource management in response to climate change and variability.
Github- daily-TWSA
Shallow water Equation Using PINNs.
In this project, I applied Physics-Informed Neural Networks (PINNs) to solve the Shallow Water Equations (SWEs), which describe the flow of fluids in oceans, rivers, and atmospheric systems. By integrating physics constraints directly into a neural network, I was able to achieve an efficient and data-driven solution to the problem.
Meesho : Predict attributes from product images.
The challenge was to develop a robust ML model to predict product attributes (like color, pattern, and sleeve length) solely from product images. Tackling this problem was an exciting experience as it addressed a critical issue in the e-commerce industry—streamlining product listing for small sellers through automation.
Github- OverfittedOrges_Meesho
Are you in safe building ?
I developed an automated ML model to identify building types and materials from street-view images. By utilizing state-of-the-art algorithms such as YOLOv8, ResNet50, and Swin Transformer, pretrained on the ImageNet dataset, we achieved a rapid and scalable approach to enhancing seismic vulnerability assessments. These models improved the accuracy of identifying building structures, contributing to better earthquake preparedness strategies.
The project was conducted as part of a Kaggle in-class competition, where I achieved the first rank. Competing with other students, I successfully applied cutting-edge ML techniques to advance the field of disaster risk management through building classification.
Page updated
Google Sites
Report abuse