Quorum

The problem

In-house counsel and entity administrators still run corporate governance out of shared drives, Word templates, and a calendar reminder for each annual filing. Resolutions get drafted from a folder of past resolutions that someone has to remember the name of. Subsidiary ownership lives in a spreadsheet nobody trusts, and filing prep is a manual checklist exercise.

Goals

• Render parent and subsidiary ownership as a live graph rather than a spreadsheet
• Draft board and shareholder resolutions from a natural-language prompt with clause-level retrieval over past resolutions
• Search and ask questions across charters, bylaws, stockholder agreements, and minutes with citations
• Run a filing-prep agent that assembles packets and pauses at a human-in-the-loop approval gate
• Enforce tenant isolation at the database row level with admin, officer, counsel, and viewer roles

The solution

• Next.js 14 app with NextAuth, multi-tenant Postgres rows, and Drizzle ORM for the entity, document, and resolution models
• Entity graph CRUD with React Flow plus Dagre auto-layout for the ownership and officer tree
• AI resolution drafting backed by Anthropic and OpenAI with a template library and clause-level RAG over past resolutions
• Document vault with pgvector embeddings, hybrid retrieval, and answer citations linking back to the source paragraph
• Realtime collaborative editor with multi-cursor presence and inline comments so counsel can redline live
• Filing-prep agent that reads entity state, gathers uploads, assembles a packet, and pauses for approval before completion
• Audit log on every state change and Langfuse traces on every model run

My role

• → Solo architect and engineer, requirements to deploy
• → Data model for entities, officers, share ledger, resolutions, documents, and the audit log in Drizzle + Postgres
• → Entity graph visualization wired to React Flow with Dagre layout and per-node detail panels
• → AI drafting pipeline: clause-level retrieval over past resolutions, template library, model selection between Anthropic and OpenAI
• → RAG vault: pgvector embeddings, chunking strategy, hybrid retrieval, citation rendering
• → Filing-prep agent with explicit human-in-the-loop checkpoint

UI direction

A counsel-facing app that looks like a modern SaaS rather than a legal tool. Tailwind, calm dark and light themes, the entity graph as the centerpiece, and a chat-style draft surface for resolutions that always shows where each clause came from.

User flows

Screenshots

Click any image to open at full size.

Close (Esc)

Key learnings

• Treating resolutions as data with structured clauses, not as opaque docx blobs, is what makes clause-level retrieval feel surgical instead of vague
• React Flow plus Dagre handles ownership trees up to a few hundred entities cleanly; beyond that, layered virtualisation is a Phase 6 problem
• An audit log table that records who, what, and when on every state change is cheaper to build on day one than to retrofit later
• Langfuse traces over every AI call (drafting, RAG, agent steps) made debugging the filing-prep agent tractable instead of mystical
• The filing-prep agent feels native because of one design choice: it pauses for human approval rather than auto-submitting anything to a regulator