Self-Improving
Legal Research

An AI research assistant that iteratively improves its legal analysis through deterministic evaluation, named mutations, and measurable score deltas.

How self-improvement works

Standard LLM

Query → Single LLM call → Answer

  • xNo retrieval from real case law
  • xCitations may be hallucinated
  • xNo way to measure answer quality
  • xOne shot -- take it or leave it
Self-Improving Pipeline

Query → Plan → Retrieve → Synthesize → EvaluateMutate → Loop

  • +RAG over 1,000+ real court decisions
  • +Citations verified against corpus
  • +5-dimension LLM-as-judge scoring
  • +Named mutations fix weak dimensions each iteration
Plan
Retrieve
Synthesize
Evaluate
Decide
Mutate
loop
This system is for legal research assistance only -- not legal advice. All outputs include citations, confidence scores, and explicit warnings when reliability is low.

Demo queries

Corpus coverage

20K+
India
SC + Delhi HC (AWS)
11K+
US
Harvard CAP + CourtListener
10K+
Canada
A2AJ / CanLII
5K+
UK
National Archives
50+
UAE
DIFC / BAILII
46K+
Real Cases
5
Jurisdictions
5
Eval Dimensions
Experimental

This is a self-improving AI research assistant that iteratively refines its legal analysis through deterministic evaluation, named mutations, and measurable score deltas. Watch the AI research, evaluate, and improve in real time.

Built by Luv Kapur