Skip to content
Abhinav Anand
work

Voice AI / Civic tech

Anvaya AI

Voice-first welfare-scheme discovery and eligibility pre-screening in Telugu, Hindi and English, with a deterministic rules engine.

Code
Anvaya AI landing page: headline 'Speak, and find the schemes that are yours', language tabs, microphone button and a live Hinglish demo card.
Landing page hero, from the built frontend.
Grafana dashboard 'Anvaya AI - Operations': catalogue size, screening outcomes, rule evaluation latency and turns by language.
Operations dashboard (Prometheus + Grafana) during a local run driven by synthetic citizen journeys, not production traffic.

[01] problem

Let a citizen find the welfare schemes they may qualify for by speaking in their own language, instead of working through portals and forms.

[02] approach

  1. 01Voice or text in Telugu, Hindi or English proposes facts, and the citizen confirms them one by one; high-impact facts such as age, state and income are read back.
  2. 02A deterministic, versioned rules engine screens the confirmed facts against 17 curated rules and returns potential or provisional matches, each with its rule trace.
  3. 03The app lists the documents to bring and prepares a draft and a copy-ready Form Helper, with an optional human reviewer behind the citizen's consent.

[03] what's interesting

  • Voice and the optional LLM are only the conversational interface; unknown facts stay unknown.
  • A kiosk mode runs the cached, checksummed rule bundle in the browser when offline and syncs queued drafts exactly once.
  • It never submits to a government portal and never reads, OCRs or stores Aadhaar numbers.

[04] stack

  • Python
  • Flask
  • React
  • Vite
  • Supabase
  • Langfuse
  • Prometheus
  • Grafana