AI / LLM / Agents
JurisGPT
Citation-grounded legal assistant for Indian startup and corporate law — RAG over 47,867 legal documents with inline citations and confidence badges. Published as an IEEE research paper.
01 / Problem
The problem
India's 100,000+ startups can't afford specialist counsel, and general-purpose LLMs hallucinate legal authority — commercial legal AI tools show 17–33% hallucination rates. Advice built on an invented section or repealed provision causes real regulatory damage.
02 / Response
The solution
Built a citation-grounded RAG assistant over a maintained corpus of 47,867 Indian legal documents (Companies Act 2013 + nine further statutes). Hybrid BM25 + lexical retrieval fused with weighted Reciprocal Rank Fusion, source-constrained generation with Claude, and every answer carrying inline citations, confidence badges, and limitation statements.
03 / Product
Key capabilities
Corpus of 47,867 Indian legal documents across 10 statutes
Hybrid BM25 + lexical retrieval with weighted RRF
Inline citation cards with expandable source passages
Calibrated confidence badges (high/medium/low/insufficient)
120-query human-verified benchmark (Fleiss' κ = 0.81)
Live corpus updates via versioned ingestion + SHA-256 manifest
04 / Outcome
Impact
05 / Reflection
What I learned
Retrieval quality in legal RAG is corpus-bound, not retriever-bound — and generic MS MARCO rerankers actively hurt legal retrieval. Rigorous evaluation with significance testing matters more than fancy components.