Deploying a fully air-gapped legal AI system inside a law firm's own infrastructure
A top-tier UK law firm (M&A and litigation practices, engagement Q4 2023 → Q1 2024) where every commercial cloud AI vendor was ruled out on day one under client confidentiality obligations, zero data leaves the firm's perimeter, period. Mistral 7B, fine-tuned on ~40,000 annotated contract clauses (4-bit GGUF), runs via vLLM on two on-premise A100 80GB servers behind an internal API. Document ACLs are enforced at the vector-store query level, a matter team only retrieves their own documents. Contract review dropped from ~6 hours to ~12 minutes. The internal IT team ran their first independent retraining cycle 3 weeks after handoff.
0
Data leaves perimeter
6hr→12min
Contract review time
100%
Team self-sufficient at handoff
The Challenge
The firm's litigation and M&A teams were spending 6+ hours per contract on initial review. Every commercial AI option involved data leaving the firm's environment, non-negotiable under their client confidentiality obligations.
The constraints were hard: no external API calls, no cloud model hosting, model must run on existing on-premise hardware, and the system must be operable by the internal IT team after handoff.
Our Approach
Discovery & Blueprint confirmed the hardware specification: two A100 80GB servers. We selected Mistral 7B with a legal domain fine-tune on 40,000 annotated contract clauses. The model runs via vLLM behind an internal API connecting to the firm's DMS. All embeddings use a locally-hosted multilingual-E5 model.
Document ACLs are enforced at the vector store query level, a matter team only retrieves their own documents.
Outcome
Initial contract review reduced from ~6 hours to ~12 minutes. Zero data leaves the firm's infrastructure. The internal IT team ran their first independent retraining cycle 3 weeks after handoff. Architecture documentation reviewed and approved by the firm's CISO.
What We Learned
01
Air-gap constraints are architectural from day one, not a retrofit.
02
Quantised open-weight models on existing on-premise hardware are production-viable.
03
Capability transfer requires deliberate design: runbooks, training, supervised handoff.
Stages Engaged
Discovery & Blueprint
Concept Validation
Production Build
Total Duration
5 months total
Artifacts Delivered
PRD
Private Infrastructure Blueprint
Fine-Tuning Specification
WBS
IT Runbook
CISO Architecture Review
Start with a Feasibility Call
2 hours. No cost. We'll tell you honestly whether AI makes sense for your case.