Built an AI-powered search and chat system enabling public access to Ministry of Statistics reports

A production-grade document intelligence and RAG chat system that lets the public ask natural-language questions across MOSPI's multi-format publications — PDFs, DOCX, PowerPoint, and CSV reports — instead of trawling through document repositories manually.

The system combines semantic retrieval with conversational responses, so a citizen, researcher, or journalist can ask a plain question (“What were the unemployment trends in 2022?”) and get an answer grounded in the original source documents.

Because the platform is public-facing, it had to hold up under concurrent, unpredictable query load while staying grounded strictly in MOSPI's own published reports — no hallucinated statistics, no answers sourced outside the ministry's documents.

I built the system end-to-end as a Django + React stack with real-time WebSocket chat via Django Channels, streaming responses from a locally hosted LLM. Voice transcription support was added so users can dictate queries.

I implemented the RAG pipeline from scratch — document ingestion with Docling OCR, embeddings into a Qdrant vector store, and semantic retrieval feeding the chat. The result is a public-facing service that operates entirely on infrastructure controlled by the Ministry.

I also handled multi-format ingestion (PDF, DOCX, PPT, CSV) so the platform could index MOSPI's existing publication library as-is, without requiring documents to be re-authored in a single format before they became searchable.