← Back to all projectsIn Development·Deployment Pending
Litigation Prep Assistant — Multi-Step RAG Pipeline
A production-style monorepo that extracts case facts, retrieves grounding statutes, drafts strategy documents, and reviews its own output end to end.
RAGMulti-Step PipelinesFull-Stack
Live demo coming soon
Technical Deep Dive
Architecture
A four-stage pipeline separates concerns that are easy to conflate into one prompt: fact extraction from raw case materials, RAG-grounded strategy generation against a statute index, drafting, and a final automated quality-review pass before anything reaches the user.
- Statute retrieval is grounded via a Pinecone vector index, so strategy output cites real statutes rather than generating plausible-sounding but unverified legal claims
- FastAPI backend and a Next.js 16 App Router frontend with Clerk authentication — structured as a deployable product with real user accounts, not a single-user notebook demo
Thought Process & Design Decisions
Splitting extraction, strategy, drafting, and review into separate stages — rather than one large prompt — makes each stage independently testable and lets the quality-review stage catch problems before they reach the drafted output, which matters more in a legal context than most.
Challenges & Trade-offs
- Balancing retrieval precision (only the most relevant statutes) against recall (not missing a statute that changes the strategy) in the RAG layer
- Designing the automated quality-review stage to catch unsupported claims without becoming a bottleneck that just repeats the drafting stage's own blind spots
Stack
FastAPINext.js 16Clerk AuthPinecone