Apparent Legal

Why I built it

Purchase agreements, LOIs, operating agreements and leases at the firm get drafted from our own precedent, which means digging through old deals for the right clause. A plain chatbot will happily draft a clause from nothing, and in legal language that's worse than useless. I wanted drafting that could only say things it could point to.

What it does

• Ingests internal precedent and classifies it by document type.
• Hybrid vector retrieval plus a reranker, with metadata filters (state, asset type, document type).
• Drafts new sections with Claude, extracting and validating a citation for every clause.
• Exposed as an MCP server, so Claude Desktop, Claude Code and team workspaces can search the corpus and draft against it directly.

How it works

Four modules: ingestion, storage, retrieval, generation. They were built on parallel branches against one shared types file that nobody was allowed to change.

Retrieval is hybrid vector search plus a reranker, with filters for state, asset type and document type. Generation drops any clause it can't tie back to a retrieved source.

Calls I made