NyayaSetu
Photograph a legal document and get it back in plain English, Kannada or Hindi, with the risky clauses flagged and the statute cited.
- Kind
- Android app and AI backend
- Year
- 2026
- Recognition
- IEEE YESIST12 2026 · Selected for the India preliminary round
- Built with
- Kotlin
- Jetpack Compose
- FastAPI
- Gemma 3
- Mistral OCR
- IndicTrans2
- 2 more
The question
What would it take for someone to understand a rental agreement, an FIR copy or a gig contract the moment they receive it?
NyayaSetu reads the document for them, in their own language, and points to the clauses and the law that matter.
Why it exists
Legal documents in India arrive in dense English that many people cannot easily read, and free legal aid exists but is hard to reach. NyayaSetu starts from the document in a person’s hand.
How it works
From a photographed document to plain language
- 1
Document
A photo or PDF of an agreement, FIR copy or contract.
- 2
Reading
Text is taken from the PDF directly where there is a text layer, and by OCR where there is not. It is then split into clauses.
- 3
Simplifying
A local Gemma 3 model rewrites each clause simply. A rewrite is rejected if its figures differ from the original.
- 4
Risk
Each clause is rated high, medium or normal. For high-risk ones the statute is fetched so the flag cites its law.
- 5
Translation
The result is translated into Kannada or Hindi and can be read aloud.
Decisions
A figure that changes is a rewrite that fails
Simplifying a legal clause must never alter an amount, a date or a percentage. The check compares the numbers in the rewrite with the source and rejects the rewrite if they differ.
Every stage can fail without losing the document
The simplifying, risk and translation stages each handle their own errors. If a model is not available the pipeline falls back to plain clause splitting, a pattern-based risk check or a simpler rewrite, so the user still gets the document back.
Cheapest reading first
A PDF with a text layer is read directly. OCR is used only for scans and photos.
Complaints go to the right forum
The consumer complaint tool chooses the District, State or National Commission from the claim value, using the thresholds of the Consumer Protection Act 2019.
Evidence
- Built
The app has screens for capture, results, question answering, RTI, consumer complaints, courts, legal aid and schemes, and a demo mode serves three cached documents (a rental agreement, an FIR copy and a gig contract).
Not built yet
- Fully tested translation: the model needs its libraries installed, and without them the app falls back to a simple word substitution
Built with
Languages
- Kotlin
Frameworks and libraries
- Jetpack Compose
- FastAPI
Data and storage
- ChromaDB
- Indian Kanoon
AI and models
- Gemma 3
- Mistral OCR
- IndicTrans2