Migrating AI decision-support to UK-based Amazon Bedrock

A GovTech business working with Cloud Combinator on AWS. The client is anonymised at their request.

The challenge had three focus areas, each one a condition for moving off the incumbent stack with confidence.

We structured the PoC as five milestones so integration, measurement and handover each had a clear gate, and so the client's engineers were brought along at every step.

The PoC delivered a working, UK-based Bedrock inference path and a repeatable evaluation framework, meeting the acceptance criteria set out at the start: functional prompt migration, an established evaluation process, end-to-end compatibility with existing configurations, UK region compliance and a validated performance measurement.

• Compatible migration — Claude 3.7 Sonnet on Bedrock replaced the OpenAI backend inside the existing inference interface, preserving the response structure that downstream Word generation relies on.
• Objective quality measurement — A Bedrock LLM-as-a-Judge framework scores outputs for Correctness and Completeness against gold-standard answers, replacing subjective checks with a repeatable process.
• Data residency by design — All inference and evaluation runs strictly within the UK region, eu-west-2, in line with the client's governance requirements.
• Performance under measurement — CloudWatch instrumentation captures average and maximum latency per document, validated against the agreed response time targets.
• Handover — A documented repository, weekly knowledge transfer and a final workshop left the client able to run future evaluations and scale the system independently.

With the PoC proving that a compliant, high-quality move to Bedrock is feasible, the natural next steps are hardening the workflow towards production, exploring more cost-effective models within the same evaluation framework, and extending coverage across the client's prompt library. These are forward-looking directions rather than delivered outcomes.