Cobalt Legal needed their junior associates to spend time reading cases, not hunting them. Their corpus: 12,000+ PDFs covering judgements, prior briefs, and statutory documents across three jurisdictions.
We built a private RAG system using hybrid retrieval (dense semantic + BM25 keyword) with a cross-encoder reranker. Citations are locked to exact source spans — if the model can't find the passage, it doesn't produce a citation.
On top of the retrieval layer, we added a memo-drafting agent. Associates prompt it with a research question; it produces a structured memo section with numbered citations. Senior counsel review and approve before anything goes out. Draft-to-review time dropped from 4 hours to under 30 minutes.
This project is in production but not publicly demoed. Details available under NDA — reach out.