ChurnSense
Production-grade user retention and churn prediction using 10K+ transaction-level records. Optimized for business impact (CLV preservation) over raw accuracy, improving retention targeting efficiency by 30%. SQL-driven EDA, 15+ RFM/behavioral features, cost-sensitive evaluation, and risk-based segmentation to minimize expected revenue loss.
AI Expense Management System
Built a role-based AI expense management system using Flask and MongoDB, enabling secure bill uploads, automated extraction, and HR approval workflows. Implemented JWT authentication, scalable backend architecture, and Railway-ready deployment for production use. Leveraged OpenAI Vision & LLMs to extract structured expense data from bill images, reducing manual processing by ~70%.
Graph-RAG arXiv
Introduced a Graph-based Retrieval Augmented Generation (Graph-RAG) system over arXiv metadata using Neo4j graph traversal instead of vector search to deliver more accurate, context-grounded answers. Established a knowledge graph schema (papers, authors, topics) and FastAPI APIs for graph retrieval, multi-hop reasoning, and LLM-driven response synthesis.
Twitter Sentiment Analysis NLP
Built a sentiment classification model (~80% accuracy) using NLP preprocessing and ML algorithms. Developed a web interface for real-time analysis via Flask API endpoints. Conducted cleaning, tokenization, stop-word removal, and vectorization.
Multi-Agent Orchestrator System with Standardized A2A Protocol
Built a multi-agent orchestration framework enabling autonomous agent-to-agent communication using a standardized A2A protocol. Designed a centralized orchestrator for task routing, context sharing, and structured message passing between specialized agents.
Incremental News Intelligence System
Built a production-grade incremental news intelligence system for continuous article ingestion, clustering, and trend detection using NLP and sentence embeddings. Implemented incremental clustering, time-decayed topic modeling, and LLM-based summaries, enabling real-time insight without retraining models. Designed a deterministic, state-persistent architecture with REST APIs and a professional analytics dashboard.
