Legal weapons

A multilingual legal AI assistant helping citizens and junior lawyers navigate India's transition from IPC to BNS, with side-by-side comparison, plain-language explanations, and Hindi/English queries.

Comment

Inspiration

It started with a simple but uncomfortable thought.

In India, there are over 50 million pending court cases. Legal help exists, but most people don’t understand how to use it. Laws are complicated. Many people, especially in rural areas, don’t know their rights, which section applies, or how to file an FIR safely.

I kept thinking:
What if people could just ask?

In their own language. In simple words. And get a clear answer.

That idea became Nyaya Sahayak , a simple legal helper.

What it does

Nyaya Sahayak lets anyone:

  • Ask legal questions in Tamil (or simple language)
  • Get clear, relevant legal sections
  • Learn their rights and available government schemes
  • Understand how to take action (like filing an FIR)

It turns complex legal text into something anyone can understand.

How we built it

We built it as a Retrieval-Augmented Generation (RAG) system:

  • Stored legal data (like BNS/IPC sections) in a vector database
  • Converted queries into embeddings
  • Retrieved the most relevant sections using semantic search
  • Improved results using reranking
  • Generated simple answers using an LLM

In simple terms:

[ \text{Answer} = \text{LLM}(\text{Query} + \text{Relevant Legal Context}) ]

We focused on making responses simple, local-language friendly, and actionable.

Challenges we ran into

  • Legal language is complex → hard to simplify without losing meaning
  • Retrieval accuracy → finding the right section is tricky
  • Multilingual support → handling Tamil queries properly
  • Data structuring → legal texts aren’t clean or uniform
  • Avoiding hallucinations → ensuring answers stay grounded in law

Accomplishments that we're proud of

  • Built a system that explains law in simple language
  • Made legal help more accessible to non-experts
  • Successfully combined retrieval + reranking + generation
  • Created a working pipeline that can map real problems to legal sections

What we learned

  • Simpler language = more real-world impact
  • Retrieval quality matters more than model size
  • Legal tech needs accuracy + clarity, not just intelligence
  • Building for real users means thinking beyond just code

What's next for Legal weapons

  • Add more regional languages
  • Improve retrieval + reranking accuracy
  • Integrate real government scheme databases
  • Add step-by-step legal guidance workflows
  • Build a mobile-friendly version for rural users

The goal is simple:
Make the law something people can actually use.

Built With

  • apache-spark-(pyspark)
  • apache-spark-pyspark
  • claude-sonnet-4
  • databricks
  • delta-lake
  • faiss
  • fastapi
  • groq
  • html
  • llama-3-70b-versatile
  • llama-3.3-70b-versatile
  • numpy
  • pandas
  • parquet
  • reportlab
  • sarvam-ai
  • scikit-learn-(tf-idf)
  • scikit-learn-tf-idf
  • sentence-transformers-(all-minilm-l6-v2)
  • sentence-transformers-all-minilm-l6-v2
  • uvicorn
Share this project:

Updates