-> What it does A multilingual legal research assistant for Indian citizens and businesses. Ask in Hindi, Tamil, Marathi (or 9 other Indian languages) — by text or voice — and get structured, citation-backed legal answers with compliance checklists, penalties, and source documents. Every answer is verified against retrieved sources before reaching the user.
-> How we built it Built entirely on Databricks — PDF ingestion → Delta table → Vector Search → Llama 3.3 70B. Added AI4Bharat IndicTrans2 for Indic↔English translation, BAAI/bge-reranker-base cross-encoder for precision reranking, an LLM-as-Judge hallucination guard, and a browser-native Web Speech API for voice input across 12 languages.
-> Challenges we ran into Databricks reranker endpoint unavailable — pivoted to local cross-encoder inference IndicTrans2 returning pandas objects instead of strings — required str() guards throughout Legal PDF chunking — built a custom heading-stack algorithm to preserve Section → Chapter → Article hierarchy Whisper pip conflicts on Databricks — switched to browser Web Speech API, which was actually faster -> Accomplishments that we're proud of Full multilingual RAG pipeline across 12 Indian languages, end-to-end Real cross-encoder reranking that demonstrably changes result ordering Two-layer hallucination guard — the answer is verified before the user ever sees it 2,325 structured chunks indexed from 22+ real Indian regulatory documents -> What we learned Retrieval quality beats prompt engineering. Unicode script detection beats language libraries for Indic scripts. LLM-as-Judge works surprisingly well when given a strict output format. And legal documents need domain-aware chunking — generic splitters destroy the structural relationships that make law meaningful.
-> What's next for Nyaya-Biz AI Expanding to GST, Labour Codes, and Companies Act. Integrating case law from High Courts. Building a WhatsApp interface so users can ask legal questions in the app they already use daily.
Built With
- ai
- ai4bharat
- apis
- bge
- databases
- delta-tables
- face
- faiss
- hugging
- indictrans2
- indictranstoolkit
- javascript
- langchain
- langdetect
- languages
- llama
- model
- models
- openai
- pandas
- platform
- platforms
- pypdf
- python
- pytorch
- re-ranker
- serving
- sql
- streamlit
- transformers
- whisper
Log in or sign up for Devpost to join the conversation.