A UK enterprise software company has built a working example of data-sovereign AI at scale. OneAdvanced, which serves more than 10,000 customers across healthcare, legal, and other regulated sectors, now runs over 50 specialized AI agents on AWS infrastructure that never leaves the United Kingdom, according to AWS’s own engineering write-up.
The motivation is regulatory, not technical fashion. Many of OneAdvanced’s customers handle patient records, legal case files, and compliance documents, and they require that no data exit the UK’s legal framework. As OneAdvanced CTO Andrew Henderson put it, data sovereignty “is a hard requirement for many of our customers, especially those in the public sector and highly regulated industries,” the company says in the same AWS post.
To meet that bar, OneAdvanced self-hosted open-weight models instead of using a managed service. It serves Llama 4 Maverick (in FP8) and Llama Guard 4 on Amazon SageMaker AI using vLLM, on p5.48xlarge GPU instances in the London (eu-west-2) region, AWS reports. The models were not yet available through UK-region managed services at the time, which is what pushed the team to run its own serving stack.
The agent layer is where the story gets interesting for builders. OneAdvanced orchestrates more than 50 task-specific agents with the open-source Strands Agents SDK, going from its first agent to over 50 in roughly three weeks, with most agents built in under a day, per the AWS account. The SDK, documented at the Strands Agents GitHub repository, takes a model-driven approach that let the team move quickly without rigid workflow definitions.
Retrieval is handled by a custom RAG pipeline the team calls “Llamadex,” built on PostgreSQL with the pgvector extension for vector similarity search. Documents are converted to markdown, chunked into 2,048-token segments, and embedded with the multilingual-e5-large-instruct model before being stored for retrieval, AWS writes.
Security and governance are the payoff. Llama Guard 4 screens inputs before they reach the main model, uploaded documents are isolated so even OneAdvanced staff cannot read them, and the company reports it holds ISO 42001 certification for AI governance and is a signatory to the EU AI Pact, according to AWS. The system has run in production since July 2025 and earned a 5/5 customer satisfaction score.
For teams weighing a similar build, the AWS post suggests starting with a two-week prototype on a managed service to validate use cases before committing to self-hosted infrastructure — a path that mirrors the one OneAdvanced itself took.
Readers exploring production agent patterns on AWS may also find value in Securing AI agents with temporal policies in Amazon Bedrock and running production AI agents in n8n with Amazon Bedrock AgentCore.
