Back to selected work

Intelligent RAG Orchestrator & Document Intelligence Engine

A document intelligence platform that ingests six content formats, retrieves relevant chunks with vector embeddings, and streams grounded answers with citations.

Period
April 2026 to present
Context
MIT Academy of Engineering
Stack
React, PostgreSQL, S3-compatible storage, JWT, Spring AI, Apache PDFBox, Apache POI, Vector embeddings

Problem

People have documents in different formats and need grounded answers that show their sources.

What I built

  • Architected a full-stack, RAG-based document intelligence platform using React, PostgreSQL, S3-compatible storage, and JWT-based authentication, with notebook-scoped document and conversation workflows.
  • Implemented ingestion for PDF, DOCX, TXT, Markdown, CSV, and web content. The pipeline extracts text with Apache PDFBox and Apache POI, creates paragraph-aware chunks of up to 500 words, and indexes them with vector embeddings for semantic search and retrieval.
  • Built streaming, source-cited RAG chat with query-intent routing and pinned-document filtering, plus a presentation editor that generates 5–8 slides from up to 20 document chunks.

Confirmed flow

RAG Orchestrator content and retrieval pipelineContent moves from upload through text extraction, paragraph-aware chunking, vector embeddings, storage, filtering, retrieval, and a streamed answer with citations. A second branch sends up to 20 chunks to a presentation editor that generates 5 to 8 slides.up to 20 chunksUploadExtract textPDFBox / POI / webParagraph-awarechunks ≤ 500 wordsVectorembeddingsStoreQuery-intent routingPinned-document filterRetrievalStreamed answerwith source citationsPresentation editorGenerate5–8 slidesRAG Orchestrator content and retrieval pipelineContent moves from upload through extraction, chunking, embeddings, storage, filtering, retrieval, and a streamed cited answer. A separate branch creates a presentation.up to 20 chunksUploadExtract textPDFBox / POI / webParagraph-aware chunksup to 500 wordsVector embeddingsStoreQuery-intent routingPinned-document filterRetrievalStreamed answerwith source citationsPresentation editorGenerate 5–8 slides
Confirmed content and retrieval flow.