Context
A boutique commercial law firm in Sydney specializing in complex commercial disputes. They needed document intelligence that fit their workflows and budget — not an enterprise solution designed for large firms.
Problem
Enterprise legal AI solutions are impressive, but they’re designed for large firms with large budgets and standardized workflows. The firm needed a system that could:
- Search across 500,000+ historical documents instantly
- Generate research reports with proper citations
- Handle discovery workflows end-to-end
- Integrate with their existing email-based workflow
- Understand Australian legal requirements and terminology
Existing tools failed because they required workflow changes, lacked Australian context, or were prohibitively expensive for a boutique practice.
What We Built
Deliverables
- Email review reports: Participant directory, email summaries, issue buckets, attachments table
- Chronology/timeline outputs: Date-ordered events extracted from the corpus
- Evidence bundle PDFs: TOC, stable document IDs, single pagination scheme
- Cited answer emails: Structured responses with source documents and date boundaries
Architecture
Tri-database retrieval system:
- Lexical: PostgreSQL full-text search for exact names, dates, parties
- Vector: Legal-specific embeddings for semantic similarity
- Graph: Neo4j for citations, relationships, document versions
Interface
Email-first. Lawyers send questions to the system and receive cited answers back — no new software to learn.
Workflow
- Input: PST exports, document attachments, plain-English questions
- Processing: Automatic indexing, entity extraction, categorization
- Review: Flagged items presented for human approval
- Output: Reports, timelines, evidence bundles
Results
- Discovery time: Reduced from 1 week to 1 afternoon for typical subpoena responses
- Meeting turnaround: Request for all communications between two parties → printed in ~10 minutes; detailed report within hours
- Citation rate: 100% — every answer includes verifiable source references
- Learning curve: Zero — lawyers started using it immediately via email
Some metrics approximated for publication.