YojnaSetu — AI Welfare Navigator for India

Multilingual AI advisor that finds every Indian government welfare scheme you qualify for Hindi, Marathi, Tamil + 5 languages. Voice input, 60 schemes, free forever.

Comment

Inspiration

India has 3,000+ government welfare schemes worth an estimated ₹1.5 lakh crore annually yet most of this money goes unclaimed. Not because the schemes don't exist. Because the system is impossibly complex to navigate.

A farmer in Rajasthan qualifies for PM-KISAN, Ayushman Bharat, Widow Pension, Sukanya Samriddhi, NREGA, crop insurance, and 6 more schemes but has never claimed a single rupee. She doesn't know she qualifies. The forms are in English. The eligibility rules are buried in government portals. Getting help requires a middleman who often demands a bribe.

We built YojnaSetu (योजना सेतु "welfare bridge") to fix this.

What it does

YojnaSetu is a multilingual AI welfare advisor that finds every government scheme a citizen qualifies for 8 Indian languages, including Hindi, Marathi, Bengali, Tamil, Telugu, Gujarati, and Kannada.

The full flow:

  1. Citizen opens the app and picks their language from 8 options
  2. Setu (our AI advisor) asks 8 simple questions: state, occupation, income, family size, land, documents, caste, and special situations
  3. The AI cross-references their profile against 60 central and state welfare schemes
  4. Results screen shows: matched schemes with confidence scores, total annual benefit in ₹, a personalized action plan in their language, and an interactive document checklist
  5. A dedicated Verify tab connects them to their nearest Common Service Centre (CSC) for human verification before applying

Key features:

  • Voice input: tap mic, speak in Hindi or any supported language, Setu understands
  • Benefits chart: visual breakdown of ₹ value per scheme
  • Text-to-speech: reads the action plan aloud for semi-literate users
  • Forward to CSC worker: sends the full plan via WhatsApp to a CSC worker who can help
  • Document checklist: interactive list with exact location, cost, and time to get each document
  • Human verification layer: confidence scoring + CSC locator + helpline 1800-121-3468

How we built it

AI layer: ** Three sequential Gemini 2.5 Flash API calls ** (1) conversational profile builder that extracts structured JSON from natural language in any Indian language, (2) eligibility matching engine that reasons about compound conditions across 60 schemes, (3) action plan generator that produces personalized Hindi/English guides.

Why Gemini beats a rules engine here: A rules-based checker would need 600+ hand-coded conditions across 60 schemes. Every policy update breaks it. It can't handle "widowed farmer with two daughters under 10 who also does construction work" a compound condition no decision tree handles. Gemini reads eligibility criteria in plain text and reasons dynamically, exactly like a human welfare advisor.

Data: 40 central government schemes from myscheme.gov.in + 20 state-specific schemes from 14 state portals, curated into a JSON database with manually verified eligibility rules, documents required, and apply links.

Frontend: React + Vite + Tailwind CSS. Deployed on Vercel. Mobile-first, works on basic Android phones. Voice input uses the built-in Web Speech API (zero cost, works in Chrome).

Challenges we ran into

  • Multilingual JSON output: Getting Gemini to return valid parseable JSON while responding in Tamil or Bengali required careful prompt engineering with strict output format instructions
  • Eligibility accuracy: Welfare scheme rules are compound and nested. Getting the AI to handle "excludes income tax payers AND government employees class A/B but NOT class D/MTS" required iteration
  • Rural accessibility: Designing for users who can't type the voice input + text-to-speech combination solved this but required careful language detection logic

Accomplishments we're proud of

  • 60 verified welfare schemes across 14 states, all with accurate eligibility rules
  • 8 Indian languages in a single app including Tamil, Telugu, and Kannada scripts
  • Voice input that works for farmers and construction workers who can't type
  • A human verification layer that CSC workers can actually use (the WhatsApp forward feature)
  • Meena's demo: a widowed farmer qualifies for 12 schemes worth ₹80,000/year in under 30 seconds

What we learned

Responsible AI in a welfare context means being wrong has real consequences a family might miss a life-changing scheme or waste time pursuing one they don't qualify for. This taught us to build confidence scoring, human handoffs, and clear disclaimers not as afterthoughts but as core product features.

What's next for YojnaSetu

  • WhatsApp bot integration (real rural reach — no app download needed)
  • Partnership with CSC Academy to train CSC workers on the tool
  • District-level state scheme expansion to all 28 states
  • Grievance filing assistant when schemes are wrongly denied
  • Real-time scheme deadline alerts (PMFBY closes June-July every year)

Built With

  • google-gemini-2.5-flash
  • myscheme.gov.in
  • node.js
  • react.js
  • speechsynthesis-api
  • tailwind-css
  • vercel
  • vite
  • web-speech-api
Share this project:

Updates