Home › Projects › SVSU Knowledge Assistant

SVSU Knowledge Assistant

Ask questions about a university's official circulars in Hindi, English or Hinglish and get answers that cite the exact page and passage. On the web and as an Android app.

  • Category: AI / RAG
  • Client: Swami Vivekanand Subharti University
  • Role: Full Stack and AI Developer
  • Status: live
  • Built with: FastAPI, Next.js 16, Flutter, PostgreSQL + pgvector, OpenAI, Tesseract OCR, Docker, GitHub Actions, AWS

Built for Swami Vivekanand Subharti University, where most notices are scanned paper circulars in Hindi and English, full of tables. Staff upload a document once: Tesseract OCR detects each page's script, rebuilds ruled tables cell by cell and flags words it was unsure of. Every paragraph is embedded and stored in PostgreSQL with pgvector. A question is rewritten into English and Hindi queries, matched by meaning plus exact codes and dates, re-ranked by a small model, and answered by GPT-4o only from those passages, with a numbered citation that opens the original scan at the cited row. A second model fact-checks the answer against its sources, and the answer streams in while it is written.

Around it: draft-to-active approval workflow, roles and per-document visibility, background bulk import, automatic expiry, two-step verification, an automatic answer-quality test, and CI/CD from GitHub Actions to an AWS server with Docker and Caddy.

Every feature is also in a Flutter Android app on the same API and accounts: answers stream in, citations open the scan at the highlighted passage, paper letters can be photographed and uploaded, and the app offers its own updates. Widget tests open every screen on a phone-sized screen in CI, and the app was tested end to end against the live server.

What it does

  • Answers in Hindi, English or Hinglish, every claim cited to the exact page and passage
  • OCR for scanned Hindi and English circulars, with ruled tables rebuilt row by row
  • Hybrid search (vectors plus exact codes, dates and rare words), re-ranking and a fact check
  • Streaming answers with live progress; the original scan opens at the cited row
  • Approval workflow, 4 roles, per-document visibility, bulk import and automatic expiry
  • Two-step verification, sign-in lock-out, and CI/CD with tests to AWS on every push
  • Android app (Flutter) with every web feature, camera upload and in-app update notices

Live site · Android app · Source code

Next project: QuickDrop