Samajh: Plain-Language Help for Indian Paperwork

The problem

A huge share of official paperwork in India arrives in dense legal-English: bank letters, electricity notices, income-tax forms, pension applications, RTI replies. Millions of people receive documents they cannot fully read, so they miss deadlines, pay charges they did not understand, or never claim benefits they are entitled to.

What Samajh does

Samajh (समझ, "understanding") turns hard paperwork into plain-language help in ten Indian languages:

• Read a Document — upload or paste a letter, notice, or form and Samajh explains it section by section in plain language.
• Listen — full text-to-speech of every explanation, for anyone who prefers hearing to reading.
• Printable checklist — a step-by-step action plan generated from the document, printable as a PDF to carry to the counter.
• Encrypted Scan Vault — scans are encrypted at rest with Fernet, using a per-user key derived with HKDF, and auto-delete after 30 days.
• Forms & Guide — step-by-step plain-language guides for 13+ common Indian forms and applications.
• My Paperwork tracker — track each document's status so nothing is lost or forgotten.
• For You & Find a Centre — relevant recommendations and nearby service centres.

How we built it

Python backend with Google's Gemma 4 (gemma-4-31b-it) through the Gemini API for explanation and translation, a plain HTML/CSS/JavaScript front end, Fernet encryption with a per-user HKDF-derived key for the scan vault, and built-in text-to-speech. All code was written during the hackathon period and is MIT-licensed at github.com/shlokgit-04/samajh.

Honest scope

• No hosted demo in this submission: the AI features call the Gemini API and need the user's own API key; the repo runs locally.
• The model only relays what our tools return. It never invents fees, documents, or URLs.
• The office-directory data shipped with the app is sample data, marked for verification on official portals.

Challenges

• Keeping the model strictly honest in a domain where a wrong fee or a wrong form number has real cost.
• Handling Devanagari and Latin scripts cleanly in one interface.
• Designing privacy for sensitive scans: per-user key derivation, encryption at rest, and automatic deletion.

Built With

• css
• encryption
• fernet
• gemini-api
• gemma
• html
• javascript
• python
• text-to-speech