Context

LexMind is a self-initiated concept product: a copilot for an in-house legal team that reviews contracts against the team's own playbooks. It was designed and built as a working demo of six connected screens, on fictional documents like the Northgate and Halden mutual NDA. The bet is that legal review is exactly the place where an assistant is most useful and least trusted, so the entire design leans into proving where every answer comes from.

Problem

A legal AI is worthless the moment it makes a claim it cannot source. Reviewers will not act on a flagged risk they cannot see in the actual paper. The design problem was provenance as the primary interaction, not a footnote: every finding has to lead straight to the exact clause in the document, in context, and every risk has to say what edit it implies and which playbook rule it fails.

Approach

LexMind is a three-pane review surface. The left holds chat history and knowledge, the middle holds the review summary, and the right holds the live document. Findings are typed edit intents with playbook rationale, and each carries a citation pill that pulses the exact clause into view when you click it. Templates are authorable schemas, and playbooks are runnable standards, so the knowledge the copilot reasons from is something the team can shape.

Outcome

Six real, connected screens where the review chats, the citations navigate, and nothing dead-ends. An NDA review flags three high risks, each traceable to its clause in one click. The product reads as a trustworthy instrument because provenance is built into the core loop rather than bolted on.