DocMind
A RAG-powered document intelligence platform — upload any document, query it in natural language, and let stateful AI agents handle multi-step reasoning workflows.
Role
Creator & AI Engineer
Capacity
Portfolio project
Year
2025–present
Stack size
10 tools
Overview
DocMind is a portfolio-grade AI engineering showcase built to demonstrate production RAG and agentic workflow practices. Users upload PDFs and documents which are chunked, embedded, and stored in pgvector. Queries are handled by a LangChain retrieval pipeline that performs semantic search over the vector index and feeds the retrieved context to an LLM to produce grounded answers. For complex, multi-step queries — summarisation chains, cross-document comparison, structured extraction — a LangGraph stateful agent takes over, maintaining conversation state and calling tools across multiple reasoning steps. The FastAPI backend exposes typed endpoints with Pydantic validation throughout. A Next.js frontend handles document management, the chat interface, and query history. The platform is containerised with Docker, deployed on AWS, and structured with a public demo repo and a private production repo.
Responsibilities
- Designed the end-to-end RAG pipeline: document ingestion, chunking strategy, embedding generation, pgvector indexing, and retrieval-augmented generation with LangChain.
- Built stateful LangGraph agents for complex multi-step workflows — summarisation, cross-document comparison, and structured data extraction — with persistent conversation state across turns.
- Implemented pgvector semantic search with tuned HNSW indexing for sub-100ms retrieval across large document corpora.
- Built the FastAPI backend with Pydantic validation throughout, typed async endpoints, and background task handling for document processing jobs.
- Developed the Next.js frontend covering document upload, chat interface with streaming responses, query history, and workspace management.
- Containerised the full stack with Docker and deployed on AWS; structured a public demo repo and a separate private production repo.
Highlights
- Full RAG pipeline from document ingestion to grounded LLM answers
- LangGraph stateful agents for multi-step reasoning — not just retrieval
- pgvector HNSW indexing for production-grade semantic search performance
- Streaming LLM responses rendered live in the chat interface