Discovery
- Worked with 6 senior associates
- Annotated 4,200 ground-truth clauses
- Mapped 38 clause types to extraction rules
AI/ML · 2024
Reading 40,000 contracts a day.
The challenge
Northwind's M&A teams were drowning in due-diligence document review. A single deal could touch 8,000 contracts. Manual review took weeks per deal.
The approach
We built a retrieval-augmented extraction pipeline tuned for legal language, with a review UI that shows reasoning beside every extracted clause.
By the numbers
0%
Extraction accuracy
0
Docs/day capacity
0%
Review time saved
0
Clause types
The process
From discovery to launch.
Architecture
- RAG over partitioned vector store
- LLM evaluation harness with quality gates
- Human-in-the-loop review queue
Design System
- Side-by-side clause + source view
- Confidence-weighted highlighting
- Reviewer keyboard shortcuts
Engineering
- Python ingest with OCR fallback
- Postgres + pgvector for embeddings
- Audit trail for every extraction
Launch
- Pilot on 3 active deals
- Calibration against partner review
- Roll-out to all M&A practice groups
Reviewer console · deal dashboard
The accuracy and traceability gave our partners confidence to actually use this in live deals. We're closing transactions faster with the same headcount.
Lawrence ParkPartner, Northwind
87%
Time saved
94%
Accuracy
40k
Docs/day
Outcomes
The numbers settled. The story didn't.
Review time per deal dropped 87% across the M&A practice. Partner adoption hit 100% within a quarter. The platform has processed 1.2M contracts since launch.
Next case study