NyayaSetu

Offline-First AI Governance Platform for Rural India

RoleFull-Stack AI Engineering, Agentic Systems Design, NLP Architecture & Academic Research

Built byShaurya Punj

LocationIndia

Year2025 – 2026

Overview

NyayaSetu (न्यायसेतु — "Bridge to Justice") is a production-grade, multilingual AI governance platform designed for rural Indian citizens and government officers with limited or no internet connectivity. It provides voice-native, Hindi/English/Hinglish access to government schemes, grievance routing, and notice management through local LLM inference — no cloud dependency required.

The Problem

Over 833 million rural Indians lack reliable internet access, yet the delivery of government welfare schemes, grievance redressal, and public notices has moved almost entirely online. Approximately 313 million adults in India cannot read or write, making text-only digital portals effectively inaccessible. Fewer than 30% of eligible beneficiaries successfully access entitlements they qualify for, due to language barriers, connectivity gaps, and complex bureaucratic interfaces.

The Solution

NyayaSetu deploys entirely on-device using Ollama-managed local LLMs (Qwen2.5-14B for structured reasoning, Sarvam-optimized models for Indic language understanding), a FAISS vector store for semantic scheme search, and Helsinki-NLP multilingual embeddings — all without requiring internet. Citizens interact via voice in Hindi, English, or Hinglish; NyayaVaani transcribes speech with faster-whisper, routes queries through a RAG pipeline, generates contextual answers, and responds via Coqui TTS — completing the full loop offline.

Project Walkthrough

Agent Architecture

NyayaVaani — Voice Intelligence Engine

Input: Spoken query in Hindi / English / Hinglish

Engine: faster-whisper → Qwen2.5-14B / Sarvam 24B → Coqui TTS

Output: Spoken answer in user's language + structured text

Speed: ~3–6s offline

RAG Scheme Intelligence Engine

Input: Natural language query (text or transcribed speech)

Engine: Helsinki-NLP embeddings → FAISS vector search → Qwen2.5-14B

Output: Grounded scheme answer with source attribution

Speed: ~1.5–4s

Grievance Router

Input: Citizen grievance text / voice submission

Engine: Qwen2.5-14B classification → SLA assignment → SQLite

Output: Categorized ticket routed to correct department

Speed: < 2s

Scheme Summarizer

Input: Raw government scheme document (PDF / text)

Engine: Qwen2.5-14B structured summarization with JSON output

Output: Citizen-friendly summary with eligibility & steps

Speed: ~5–10s

Notice Drafter

Input: Officer intent / template + auto-filled credentials

Engine: Qwen2.5-14B generation with role context injection

Output: Formal government notice draft ready for review

Speed: ~3–7s

System Architecture

Technology Stack

client

React 18
TypeScript
Electron (Desktop)
HashRouter (Offline SPA)

server

FastAPI (Python)
SQLite (Offline-First)
REST API + OpenAPI Docs

ai

Ollama (Local LLM Runtime)
Qwen2.5-14B
Sarvam 24B (Indic Intelligence)
faster-whisper (Offline ASR)
Coqui TTS (Offline Synthesis)
FAISS Vector Search

data

SQLite (All Persistence)
FAISS Vector Index
Government Scheme PDFs
In-Memory Cache (280x Speedup)

Key Features

Fully offline — all LLM inference runs locally via Ollama, zero cloud dependency

Voice-native multilingual interface: Hindi, English & Hinglish supported

RAG-powered scheme discovery grounded in real government documents

Role-based access: citizen self-service vs officer admin panel

Grievance routing with department assignment and SLA monitoring

AI-assisted notice drafting with officer credential auto-fill

Screenshots