Inside OneAdvanced's UK-sovereign agent platform
OneAdvanced has built a UK-sovereign AI platform on AWS by self-hosting Llama 4 models in the London region, according to a technical case study published by AWS. The system combines SageMaker AI endpoints, more than 50 task-specific agents on Amazon ECS, and a retrieval pipeline backed by pgvector. The deployment is aimed at regulated customers that need AI features without moving user data outside the UK.Sovereignty drove the decision to self-host models
AWS says OneAdvanced, a UK enterprise software provider serving more than 10,000 customers, needed AI capabilities for sectors including healthcare, legal and other regulated industries. The constraint was not only model performance. Customers handling patient records, legal case files and compliance documents wanted assurance about where data resided and who could access it.OneAdvanced initially prototyped with Amazon Bedrock during a two-week sprint, including chat completion, an agent for querying UK statute law, Snowflake integration and chart generation. The company then chose a self-hosted route because the specific models it wanted - Llama 4 Maverick and Llama Guard 4 - were not yet available through AWS managed services in the UK region at the time of the engagement.
That decision shifted responsibility to OneAdvanced and AWS infrastructure under its control. Instead of relying on a managed model service, the company had to deploy, serve and scale open-weight models while adding moderation, document retrieval, orchestration and a no-code agent builder around them.
The model stack runs on SageMaker in London
The deployment uses vLLM to serve Llama 4 Maverick and Llama Guard 4 on Amazon SageMaker AI endpoints, according to AWS. The models run on p5.48xlarge instances in the London, eu-west-2, region, using Hugging Face models with AWS Deep Learning Containers.AWS says OneAdvanced moved from p4d.24xlarge instances to p5.48xlarge instances for production because of longer context-window needs. The company targets 120K to 128K token context lengths for workloads such as large document analysis and multi-turn conversations. Load testing with vLLM on P5 instances was used during the advisory engagement to validate throughput requirements.
Llama Guard 4 is positioned before the main model in the request path. AWS says it replaced an earlier Llama Guard 3 deployment after OneAdvanced observed high false rejection rates with the previous version. The guard model screens user inputs for harmful content before inference begins.
More than 50 agents sit above the model layer
A central feature of the platform is an agent catalog built with Strands Agents SDK. AWS says OneAdvanced went from its first agent to more than 50 in three weeks, with most agents built in less than a day.The agents cover sectors and functions including healthcare, legal, HR, marketing, logistics and education. Examples cited by AWS include a Care Incident Response assistant, a Clinical Safety Bulletin generator, a scheme of work generator, a Document Comparison tool and an AWS Architect Agent. Each agent has a system prompt, selected tools and an optional structured input form.
The runtime design separates agent behavior from infrastructure. Agents are containerized on Amazon ECS, while runtime configuration is stored in Amazon DynamoDB. Users select agents from a catalog, and non-developers can create new agents through a visual no-code builder that defines personas, form fields, prompts and tool access.
Retrieval uses S3, Aurora PostgreSQL and pgvector
The platform grounds agent responses in uploaded documents through a retrieval augmented generation pipeline. Documents are stored in Amazon S3, converted to markdown, chunked into 2,048-token segments and embedded into pgvector for similarity search.AWS says the database layer is Amazon Aurora PostgreSQL-Compatible Edition with the pgvector extension. OneAdvanced uses the intfloat/multilingual-e5-large-instruct embedding model, selected for multilingual support and instruction-following capabilities. The company built its own retrieval system, internally called Llamadex, rather than adopting an off-the-shelf RAG framework.
Retrieval is exposed to agents through tools such as Personal knowledge search and Organization knowledge search. When an agent uses uploaded content, users can see which documents were used to generate a response. That transparency matters in regulated environments where a generated answer is less useful if the source trail is hidden.
Controls are framed around data residency and governance
AWS says the infrastructure runs in the London AWS Region and that no user data leaves the UK. It also says user queries and responses are not retained or used for model training, and that uploaded documents in personal and organization spaces are isolated and not readable by OneAdvanced staff.The platform includes several control layers described in the AWS case study. Llama Guard 4 screens harmful inputs before the main model. Amazon GuardDuty provides threat detection for the S3 document storage layer. Privacy controls can be customized at organizational levels.
OneAdvanced holds ISO 42001 certification for AI governance, according to the source, and is a signatory to the EU AI Pact. Those claims do not by themselves prove a model is safe or reliable, but they show the company is presenting governance as part of the platform architecture rather than as an afterthought.
Conclusion
The OneAdvanced deployment is a useful example of how sovereign AI requirements can change a cloud architecture. The main story is not just that the company built many agents, but that it did so while self-hosting selected Llama 4 models in a specific UK region and adding retrieval, moderation and governance controls around them.The evidence comes from an AWS-published case study co-authored with OneAdvanced, so claims about market position and launch status should be read as company-reported rather than independently verified. The technical pattern is still notable: regulated AI deployments may increasingly depend on where models run, how documents are retrieved, and how agent tools are governed.
Sources
Editorial Team - CoinBotLab
