Girish
Jeswani

‹ ›

GradCompass is an AI-powered graduate application assistant built with React 19, FastAPI, and Google Gemini AI that guides students through their entire grad school journey via 5 specialized AI agents. The platform features real-time WebSocket-powered mock visa interviews, intelligent university matching, and document assistance — all backed by async PostgreSQL operations and LangGraph workflow orchestration. Think of it as having a personal admissions counselor who never sleeps, powered by modern full-stack architecture that makes the overwhelming world of graduate applications actually manageable.

‹ ›

Cine-Stellation is an interactive movie recommendation system that visualizes movie relationships as constellation graphs using machine learning algorithms. Built with Next.js frontend and FastAPI backend, it leverages TF-IDF vectorization and cosine similarity to identify movie connections based on content and user preferences. The system features real-time plot-based search, genre filtering, user authentication with MongoDB persistence, and force-directed graph visualizations rendered on HTML5 Canvas. Users can explore movie recommendations through an immersive space-themed interface while tracking their personal watchlists and discovering new films through intelligent similarity matching.

‹ ›

Talk to Your Repo is an AI-powered code assistant that enables natural language conversations with GitHub repositories using a modern full-stack architecture. Built with FastAPI (Python backend) and React (frontend), it leverages Google Gemini AI for advanced code embeddings and semantic understanding, automatically cloning and indexing repositories into searchable chunks. Users can ask questions about code functionality, structure, and implementation through an intuitive chat interface, receiving intelligent responses with precise source citations, file references, and real-time processing status updates. The system features vector similarity search for relevant code discovery, file tree visualization, conversation history, and intelligent filtering that processes only relevant code files while maintaining context across multiple queries.

This project developed an integrated wildfire risk prediction system for Queensland, Australia using multi-source data from NASA FIRMS (satellite fire detection) and Open Meteo weather APIs covering 2021–2024. The team implemented and compared seven machine learning models including XGBoost, LSTM, Random Forest, and SVM using features like temperature, humidity, soil conditions, wind speed, and precipitation. LSTM achieved the best performance with 82% accuracy and 90% recall, while XGBoost delivered strong balanced metrics at 80% accuracy. The system is designed for real-time wildfire occurrence prediction and early warning applications.

This project develops a Visual Question Answering system that combines Stacked Attention Networks with Multi-Layer Feature Fusion to improve image understanding. The system uses VGG-19 for image feature extraction and LSTM for question processing, but introduces a novel approach of fusing features from multiple CNN layers (rather than just the final layer) using addition and concatenation operators. The stacked attention mechanism helps reduce noise and focus on image regions relevant to the question, while multi-layer feature fusion preserves important information that would otherwise be lost in deeper layers. The approach achieves 63.72% accuracy on the VQA dataset, outperforming baseline methods including vanilla VQA (57.58%) and single attention mechanisms (59.8%).

This project develops a mobile wallet payment application using React Native, Node.js, and MongoDB that enables secure transactions in both online and offline environments. The system uses AsyncStorage for offline transaction storage, WebSockets for secure online communication, and bcrypt for password encryption, addressing connectivity issues in areas with poor internet access. Key technical features include document-level database locking for concurrency control and automatic synchronization of offline transactions when connectivity returns. Performance testing showed 600ms average transaction times, outperforming existing wallet applications that typically require 700–800ms. Published in SN Computer Science (2022).