Saudia's Digital Transformation: Data Center Migration to Google Cloud Platform in Damma

We helped a secure government agency deploy an air-gapped AI search engine that understands Arabic morphology without a single byte of data leaving their facility.

Government Ministry The Client
99% Faster Discovery The Win
Bunyan + Local LLM The Tech
12 Weeks Time to Value

The Problem

The Problem

The Ministry’s legal team manages a repository of over 10 million scanned PDF pages dating back to the 1970s. Their existing keyword search infrastructure was failing to capture legal nuance, forcing paralegals to spend up to 48 hours manually cross-referencing physical files for a single case.

Standard cloud-based RAG (Retrieval-Augmented Generation) solutions were rejected due to three hard boundaries:

  • Zero internet access: The entire system had to run offline.

  • Strict classification: A Junior Lawyer cannot see a Senior Partner’s files.

  • The Language: Standard English tools failed to understand the nuance of legal Arabic.

The Solution Architecture

We engineered a secure intelligence pipeline that runs entirely on-premise.

Core Components:

  1. Document Ingestion: A custom OCR pipeline that processes scanned PDFs, corrects character errors, and structures the text for analysis.

  2. Vector Database: A self-hosted index (Milvus) that stores data as "concepts" rather than keywords, enabling semantic retrieval.

  3. Local Inference Engine: A quantized 70B parameter Large Language Model (LLM) hosted on local H100 GPUs, tuned for legal reasoning.

  4. Access Control Layer: A security gateway that cross-references every query against the user's Active Directory permissions before returning results.

Technical Specifications

Model

Llama-3-70B (Quantized 4-bit)

Compute

Compute

Compute

Compute

Compute

Compute

Implementation Timeline

  1. Month 1: Data Audit. We mapped the document classification levels and validated the hardware specs for local inference.

  2. Month 2: Deployment. We installed the Bunyan platform and trained the embedding model on the specific legal corpus.

  3. Month 3: Hardening. We conducted Red Team testing to ensure that classified documents remained invisible to unauthorized users.

Key Technical Challenge: The standard model tokenizer fragmented Arabic legal ligatures, resulting in poor retrieval density.

We rebuilt the tokenizer vocabulary using the Ministry’s specific document corpus, improving semantic relevance scores by 40% without increasing inference latency.

Operational Outcomes

The transition allowed the legal team to move from manual archiving to instant retrieval.

Metric Legacy Process OmniOps Deployment
Search Latency 48 Hours (Manual Review) 48 Hours (Manual Review)
Search Precision Keyword Match Only 92% Semantic Relevance

OmniOps didn't just drop a model and leave. They engineered the RBAC controls so our CISO signed off on Day 1. It is the only AI system currently approved for our classified network.

Director of Legal Operations [Government Ministry]

Current Operations

The Ministry’s IT team now fully owns the platform. OmniOps provides model updates and support, but the client retains full control over the hardware, the data, and the access logs.