ShehriSaathi (شہری ساتھی, citizen's companion)

Ask civic questions by voice in Urdu and get simple, source-cited answers drawn only from verified government documents. Tool: https://shehri-saathi.vercel.app/

Comment 1

ShehriSaathi — Citizen's Companion

Millions of Pakistanis can't easily access clear civic information. Voter registration steps, how to get a CNIC, and how to file a complaint are scattered across government websites in dense English or formal Urdu — unusable for someone with low literacy, on a low-end phone, with limited data. ShehriSaathi lets a citizen simply ask a question by voice in Urdu and get a short, clear answer — drawn only from official documents, with the source cited every time.

What it does

  • Accepts questions by voice or text in Urdu.
  • Answers in simple, clear Urdu suitable for low-literacy readers.
  • Answers only from a curated set of official Pakistani civic documents (Election Commission of Pakistan, NADRA, Pakistan Citizen Portal) — and cites the source of every answer.
  • Refuses to guess. If a question falls outside its verified documents, it says so in Urdu instead of inventing an answer — directly tackling the misinformation problem.

Inspiration

Standard AI chatbots fail exactly the people who need civic help most: they hallucinate legal and procedural details, assume English literacy, and assume fast connections. We wanted the opposite — a tool that is honest about what it knows, speaks the user's language by voice, and runs light enough for the realities of a low-bandwidth, low-end-device user.

How we built it

  • Frontend & server: Next.js (App Router) + TypeScript, RTL Urdu UI with the Noto Nastaliq Urdu font, mobile-first.
  • Language model: Llama 3.3 70B served on Groq's fast inference (free tier, OpenAI-compatible endpoint).
  • Voice input: Groq-hosted Whisper (large-v3-turbo) for Urdu speech-to-text.
  • Grounding (RAG): Official content is ingested into local markdown documents, chunked, and retrieved with an in-memory BM25 keyword index. The model is constrained to answer only from the top retrieved chunks and to cite them.
  • Automated ingestion pipeline: A re-runnable script fetches content from official government sources and writes it to the document store, with a committed cached fallback so the app always works even offline.

Why AI is the right tool

The problem isn't a lack of information — it's that the information is unreadable for the target user. A language model is uniquely able to take dense official text and restate it in simple spoken Urdu, on demand, for any phrasing of a question. Pairing it with retrieval and forced citation keeps it accurate where a plain chatbot would hallucinate.

Challenges we ran into

  • Preventing hallucination of civic facts, where a wrong answer is worse than none — solved with strict retrieval grounding, source citations, and an explicit Urdu refusal when context is missing.
  • Urdu rendering and RTL layout with a readable Nastaliq font on small screens.
  • Working within free-tier limits — we kept retrieval lean (top-4 chunks) to stay inside the token budget.
  • Unreliable access to government sites — handled with a cached document fallback so a demo never depends on a live site.

What we learned

For civic tools, trustworthiness beats fluency. A model that admits "I don't have verified information on that" is more valuable to a citizen than one that sounds confident and is wrong.

What's next

  • Replace the curated document set with direct connections to official open-data feeds so it auto-updates.
  • Add more regional languages and full voice-out answers.
  • Pilot delivery over WhatsApp to reach low-end phones with no app install.

Built With

Share this project:

Updates