Flo Health Details a Bedrock Pipeline for Medical Content Review

Flo Health and AWS described a production workflow that uses Amazon Bedrock to help review medical content while retaining final human approval. The system connects a Contentful publishing workflow through API Gateway, S3, Lambda, Step Functions, and DynamoDB. It applies separate model-based judges for medical accuracy, legal compliance, and brand style rather than asking one prompt to handle every review dimension.

Flo says lighter review work is routed to Haiku and more complex tasks to Sonnet. The companies report 60% lower review time, threefold throughput without expanding the medical team, more than 70% fewer repeated errors, and automation of 80% of routine compliance corrections. These figures come from the coauthored company case study and have not been independently audited. Human medical reviewers remain the final approvers.

The design separates orchestration, content storage, model calls, judge outputs, and workflow state. That separation matters because medical accuracy, legal compliance, and brand style have different evidence requirements and different consequences when a model misses an issue. Routing by task complexity can control cost, but only if the router itself is measured for false negatives and escalation failures.

A modular judge architecture is easier to audit than one general quality score, but independence between judges should not be assumed merely because prompts are separate. Teams should test correlated failures, use adversarial examples, measure agreement with qualified reviewers, and maintain a holdout set that changes only through controlled review.

The workflow also needs evidence lineage. Every suggestion should retain the exact content revision, cited medical source, model and prompt version, judge outputs, policy version, reviewer decision, and final published change. If a source is retracted or a policy changes, the organization should be able to identify affected content and rerun the relevant checks.

LDS views the strongest design choice as separating judges and preserving human approval. The main unresolved risk is whether the system measures the errors it does not catch. Throughput and correction rates show operational efficiency, but they do not prove medical correctness. A mature deployment should publish internal quality thresholds for false negatives, reviewer overrides, disagreement, and post-publication corrections, then halt automation when those thresholds regress.

Useful follow-up evidence would include independent audit results, judge-level precision and recall, error severity distributions, reviewer-override rates, correction latency, and proof that the workflow fails closed when sources, models, or policy checks are unavailable.