Saathi
Bilingual Hindi/English AI assistant that checks welfare scheme eligibility and prepares a citizen-verified document checklist.
Built at AgentX Hackathon 2026
Team lastmile coder: Krunal Chikhale, Om Dubey
Built with Kotlin, Jetpack Compose, Room, Next.js, React, TypeScript, Tailwind CSS, Node.js, Express, PostgreSQL, SQLite, Firebase Auth, Firebase App Check, Gemini AI, Docker, Playwright
About the project
The problem: Eligible citizens often miss government welfare benefits because scheme rules (age cutoffs, state-specific income limits, landholding exclusions) are buried in dense official portals. One missing document or a misread rule leads to rejections and repeated visits to local centres.
Our solution: Scheme Saathi is a bilingual (Hindi/English) assistant that tells citizens which schemes they qualify for and which documents they need, before they go to the portal or a CSC centre.
How it works: A deterministic eligibility engine evaluates ALL/ANY rule trees across 5 schemes (PM-KISAN, NSAP-IGNOAPS, PMUY, PM-JAY, APY), with state overrides and an UNKNOWN status when data is missing. Users can talk to the AI by voice. For example, a spoken age like "बासठ" is echoed back as 62 for confirmation. A Human Confirmation Gate requires the citizen to review their details and generates a SHA-256 snapshot hash before a CSC Guided Handoff Pack unlocks. The AI never auto-submits anything.
Safety and privacy: 12-digit Aadhaar numbers are rejected, PII is redacted before any LLM call, profile fields use AES-256-GCM encryption, and role-based access (Citizen, CSC Operator, Reviewer, Supervisor) is enforced on the server. DPDP-style consent, data export and erasure are included.
Impact: Fewer rejected applications, less dependence on middlemen, and easier access for Hindi speakers and low-literacy users, with Large Text mode and voice input. One shared backend serves both the Android app and the web PWA.