Hackathon Submission – E‑Mitra (ई‑मित्र)

Inspiration

• Migrant workers in Delhi face fragmented information about labour rights, welfare schemes, and local offices.
  
• The COVID‑19 pandemic highlighted the need for a trusted, always‑available digital assistant that can speak the worker’s language and guide them through complex regulations.
• We wanted to empower workers with instant, conversational access to legal statutes, wage rules, and nearby labour offices.

What it does

• Multilingual conversational assistant (English + Hindi) that answers labour‑law queries using a Retrieval‑Augmented Generation (RAG) pipeline.
• Geolocation‑aware routing – when location is allowed, it auto‑detects the nearest district office; otherwise it gracefully asks the user for their district.
• Voice‑first UI – press‑to‑talk for hands‑free interaction, with real‑time transcription.
• Secure API – JWT authentication, rate‑limiting, Helmet headers, and Zod validation protect the backend.
• Offline fallback – static fixtures provide answers when Elasticsearch is unavailable.

How we built it

• Frontend: React (Vite) with a minimalist UI, dark/light mode toggle, and a voice recorder component. State is managed with React hooks; API calls use fetch with a conditional payload to avoid sending null values.
• Backend: Node / Express server, JWT‑protected routes, Zod schemas for strict request validation, and express-rate-limit throttling.
• Search & RAG: Elasticsearch stores statutes, schemes, and district‑office data. Vectors are generated via Gemini embeddings; the chat route builds a context‑rich prompt and asks Gemini to return a JSON‑structured answer.
• LLM integration: Gemini is used for language detection, translation, embedding generation, and final answer synthesis.
• Security: Helmet CSP, compression, input sanitisation, and audit logging (fire‑and‑forget) ensure data privacy and resilience.

Challenges we ran into

• Zod schema mismatches when the frontend sent { lat: null, lon: null }; fixed by sending only defined fields.
• Geolocation denial – required a conversational fallback that still respects privacy while prompting the user for their district.
• Prompt brittleness – Gemini sometimes wrapped JSON in backticks; added robust parsing and a clean‑up step.
• Multilingual consistency – ensuring Hindi answers used common, worker‑friendly terminology rather than formal Sanskritized Hindi.

Accomplishments we’re proud of

• End‑to‑end voice‑enabled chat that seamlessly switches between English and Hindi.
• Dynamic office routing that works with both GPS coordinates and user‑provided district names.
• Hardened security stack (CORS, Helmet, rate limiting, JWT) while keeping the demo lightweight for a hackathon.
• Fully responsive UI with dark/light mode toggle and a polished micro‑animation for the microphone button.

What we learned

• LLM‑centric design: prompting strategies and strict JSON schemas are essential for reliable integration.
• Security‑first mindset: even in a demo, applying OWASP best practices avoids needless vulnerabilities.
• User‑centric fallback flows – gracefully handling denied permissions preserves trust and improves experience.
• Rapid prototyping: combining Vite, Elasticsearch, and Gemini allowed us to iterate within days.

What’s next for E‑Mitra

• Expand the knowledge base to cover other Indian states and central labour laws.
• Add offline caching of recent queries for low‑connectivity scenarios.
• Introduce a progressive web‑app (PWA) install flow for easy access on feature phones.
• Refine district‑detection with fuzzy matching and ML‑based location inference.
• Open‑source the project and encourage community contributions via a public GitHub repo.

Built With

• elasticsearch
• node.js
• python
• vite