Grounded answers, never guesses

A Healthcare & Life Sciences business working with Cloud Combinator on AWS. The client is anonymised at their request.

The challenge had four focus areas, each shaping the architecture rather than sitting on a checklist.

The engagement runs as one fixed-scope phase over four to six weeks, structured so that the accuracy target is measured rather than asserted, with production scoped separately afterwards.

Acceptance is defined against measurable, grounded behaviour rather than a subjective demo. The system is accepted when the retrieval layer answers questions over the graph on AWS, meets the agreed accuracy target and makes every answer traceable. The figures below are the agreed targets and design guarantees set out in the Statement of Work.

• Measured accuracy — A dedicated evaluation harness scores responses against a mutually agreed set of representative questions, so the accuracy target is proven, not claimed. The proposed floor is 80 percent, chosen to give the delivery team headroom to over-deliver.
• Grounded or abstain — Schema-grounded generation with live schema validation means the agent returns a result it can stand behind or explicitly declines, removing the fabrication that disqualified the earlier approach.
• Full traceability — Every answer carries the query that produced it, so any result can be checked back to source, which matters for a regulated research team.
• Controlled access in-region — The tool runs in eu-west-2 (London) and is reachable only by authenticated the client's staff through Amazon Cognito federated to their identity provider, with unauthenticated requests blocked at CloudFront and WAF.
• Handover — The engagement closes with an evaluation report, the confirmed architecture and a production recommendation, delivered through the collaborative, hands-on model the client asked for.

With grounding, measured accuracy and portability designed in from the start, the client finishes the proof of concept with a retrieval layer they can trust and a clear, reusable path into a production build, ready to put a trustworthy plain-English window over their knowledge graph and to extend it as their research grows.