Projects
A showcase of my recent work, featuring web applications, AI tools, and experiments that demonstrate my technical skills and creativity.
Web Development Projects
Featured Projects
Production-grade multi-tenant CRM platform engineered to handle 1,000+ concurrent users with Bull queue offloading delivering 3x faster response times than naive REST approaches. Features org-level RBAC, JWT auth with 2FA, and org-scoped Socket.io rooms for isolated real-time chat. Implemented Cloudinary file handling, webhook support, lead scoring, deal pipeline tracking, analytics dashboard with CSV/Excel export, per-org rate-limited AI assistant (Nemotron) with streaming inference, and real-time notifications.
Razorpay-integrated credit-economy platform where posting costs credits and earning upvotes rewards contributors — creating self-moderating content quality loops. Built with Clerk + JWT auth, real-time Socket.IO messaging, and PostgreSQL optimized for high concurrency. Includes full moderation pipeline, media-rich posts/comments, user profiles, and scalable RESTful APIs.
Other Web Projects
Socket.IO-powered social platform with live messaging and real-time notifications across concurrent users. Built on MERN stack with JWT auth, Cloudinary media uploads, and modular service architecture for horizontal scalability. Supports post reactions, nested comments, full-text search, and advanced security — rate limiting, input sanitization, and CORS policies.
Stripe-integrated full-stack LMS with Redis caching for high-throughput query performance, serving students and instructors with role-based access control. Built modular architecture with real-time messaging, JWT auth, and RESTful APIs. Optimized MongoDB with strategic indexing for fast course content delivery at scale.
Weather dashboard built with EJS, Node.js, MySQL, and OpenWeather API. Provides real-time weather updates, forecasts, and historical data visualization. Features location search, favorite cities, and customizable themes. Backend includes scheduled data refresh and caching for performance.
AI & Machine Learning Projects
Featured Projects
Multi-agent RAG platform that unifies enterprise knowledge across Gmail, Drive, Notion, Slack, and PDFs into a single intelligent interface. A LangGraph supervisor routes queries to 11 specialized agents, retrieves context via hybrid vector (Qdrant) + graph (Neo4j) search, detects contradictions with DeBERTa NLI, and returns cited confidence-scored answers. Implements 6 RAG patterns: Naive, Retrieve-and-Rerank (CrossEncoder top-20→5), Graph RAG, Hybrid RAG, Agentic RAG, and Multi-Agent RAG.
Emotion classifier achieving 88% accuracy across 27 emotion classes on the GoEmotions dataset (58,000+ Reddit samples). Built separate transformer and LSTM models — fine-tuned DistilBERT for state-of-the-art results and implemented an LSTM baseline for comparative analysis. Includes full data preprocessing pipeline, model evaluation, and confusion matrix visualization. Explored transfer learning and ensemble methods for further accuracy improvements.
Other AI Projects
Built a custom multitask transformer from scratch in PyTorch with multi-head self-attention and cross-attention, using a shared encoder with per-task decoder heads for a unified text-to-text framework. Trained on translation, summarisation, and question-answering using task-conditioned prompts and masking strategies. Evaluated with BLEU, ROUGE, and Exact Match metrics.
A Retrieval Augmented Generation (RAG) system for question answering over PDF documents. Users can upload a PDF, which is split into semantic chunks and embedded using sentence-transformers. A FAISS vector store enables fast similarity search, and relevant chunks are retrieved for each query. The DeepSeek LLM generates answers using both the user's question and the retrieved context, providing accurate, context-aware responses. Built with LangChain, Hugging Face Transformers, ChromaDB, and FAISS.
Content-based filtering for personalized movie recommendations. Utilizes user ratings and movie metadata to generate tailored suggestions. Implemented both TF-IDF and Count Vectorization methods for feature extraction and similarity scoring. Built with Python, Pandas, and Scikit-learn, and supports scalable batch processing.
Spam detection using machine learning algorithms. Trained on large datasets with feature extraction using NLTK and Scikit-learn. Implements multiple classification models, including logistic regression and random forest, with cross-validation for accuracy. Includes a web interface for testing and visualizing results.