Built an agentic RAG chatbot

Built an agentic RAG chatbot enabling natural language conversations over 25 landmark AI/ML research papers from ArXiv, with source-attributed answers and page-level citations.

Designed a 5-category intent classification system using zero-temperature Gemini to route messages - eliminating unnecessary vector.

Implemented hybrid retrieval combining FAISS vector search (60%) and BM25 keyword search (40%) using LangChain EnsembleRetriever, improving recall for both semantic queries and exact terminology.

Architected a two-layer memory system - short-term session memory and long-term cross-session memory, injected into the system prompt for personalised responses.

Deployed FastAPI backend on AWS EC2 with Ngrok HTTPS tunnel and React frontend on Vercel.

Technologies: Python, FastAPI, LangChain, HuggingFace Transformers, FAISS, BM25, Google Gemini, SQLite, React, AWS EC2, Vercel.