Built a document intelligence system processing decades of historical government reports for 1000+ users

A full-stack document intelligence platform — the Feeder Platform — for the Ministry of Statistics, serving 1000+ users with automated analysis and processing of decades of historical government reports.

Where the public-facing search tool helps citizens find answers, the Feeder Platform is the internal tool that creates the knowledge base: it OCRs, extracts tables, and converts unstructured legacy documents into queryable data that downstream platforms can use.

Government statistical archives accumulate decades of reports in inconsistent formats — scanned tables, mixed layouts, non-standard headers — that make bulk automated processing genuinely hard. The pipeline had to handle that heterogeneity without a human manually reviewing every document, while still producing structured output reliable enough to power natural-language SQL queries downstream.

I designed and built the AI/ML pipeline using Docling, Vanna.AI, and Ollama for automated OCR, table extraction, and natural-language-to-SQL conversion — reducing document processing time by 80% compared to manual workflows.

I integrated a Qdrant vector database for semantic search and RAG-based chat, and deployed the stack on an H200 GPU server running an in-house OpenAI OSS 120B model — so sensitive ministry data never leaves the Government's infrastructure.

I also built the full-stack application layer end-to-end, giving ministry staff a single interface to upload legacy documents, monitor extraction status, correct OCR/table-extraction errors where needed, and query the resulting structured data in plain English instead of writing SQL by hand.