Personalised, evidence-grounded gut health reports

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

The challenge had three focus areas.

The pipeline uses retrieval-augmented generation so that synthesis is always tied to retrieved evidence. An Amazon API Gateway endpoint accepts a request containing the user's microbiome profile and metadata and routes it to an AWS Lambda function. The function benchmarks the profile against the healthy reference cohort and then calls the Amazon Bedrock agent runtime Retrieve and Generate API, which queries the Project 2 knowledge base, combining graph traversal and vector similarity in a single call, retrieves the relevant literature, and generates the report sections with inline citations.

The engagement delivered an end-to-end AI reporting platform: a RAG pipeline, a serverless report-generation API, an engineered prompt library, an automatic citation engine with grounding-confidence scores, a SME review workflow for prompt iteration, full infrastructure as code, and documentation. It gives the client a way to turn a microbiome profile into a trustworthy, cited report on demand. Figures below reflect the design targets set in the scope of work.

• Evidence-grounded by design — A citation engine injects source references and grounding-confidence scores automatically, linking each claim in the report back to the papers behind it.
• Retrieval plus synthesis in one call — The Bedrock Retrieve and Generate API combines graph traversal and vector search against the knowledge graph before generating text, keeping output anchored to retrieved evidence.
• Cost-matched models — Haiku handles high-volume retrieval while Sonnet produces the final narrative, balancing quality against the cost of running at scale.
• Safety built in — Bedrock Guardrails apply content filtering, topic restrictions, and PII detection to every generated report.
• Handover — Cloud Combinator delivered the platform as infrastructure as code with a prompt library, API specifications, a SME review workflow, and operational runbooks, so the client can refine prompts and operate the service.

With the reporting prototype completing the three-project engagement, the client has a full path from raw literature to a personalised, cited report, and a serverless foundation it can iterate on with its subject-matter experts as prompt quality and grounding accuracy are validated in testing.