This is an anonymised design account, revised against current AWS and privacy guidance. It describes a reference architecture intended to support privacy controls; it does not establish production traffic, performance, cost, or legal acceptance.
Keep the request path narrow
The useful design question is where a customer message can travel, and who can retrieve it later. The reference path is deliberately explicit: authenticate the request, apply purpose and data-minimisation rules in the application layer, invoke the selected model in an approved Region, then write only the records that the product needs. Structured state can live in DynamoDB; selected objects can live in S3; KMS manages the encryption boundary; CloudTrail and CloudWatch provide operational evidence.
That sequence is an implementation choice, not a claim that a service name creates compliance. The Australian Privacy Principles quick reference describes APP 3 and APP 6 obligations around collection and later use or disclosure, APP 8 cross-border disclosure, and APP 11 reasonable security and handling when information is no longer needed. Each obligation still needs to be assessed against the actual data flow. A New Zealand deployment should be reviewed against the New Zealand privacy principles on its own terms.
request -> authenticate -> minimise for purpose -> model in approved Region
-> return response -> store approved fields -> record access event
This is illustrative pseudocode, not executable SDK code or a test result. It makes the missing decisions visible: the purpose, the approved model path, the retention policy, and the identity attached to each access.
Treat the model boundary as a setting
Amazon Bedrock is a useful managed-model boundary for this design, but “Bedrock” is not a retention policy. AWS documents different retention modes, including none, default, and aws_review; their effect depends on the account, Region, project, model path, and configuration. Before making a privacy statement, verify the setting for the exact route and check whether cross-Region inference or provider terms change where data is processed.
The same rule applies to the Sydney Region (ap-southeast-2). Choosing an Australian Region can be part of a data-location strategy when each selected service supports it. It does not prove that logs, backups, support paths, model inference, or every dependent service remain there. The control is the verified end-to-end path, with an exception process for anything that crosses it.
Put storage and key ownership on a reviewable boundary
The design separates short-lived conversational state from durable objects and makes retention an explicit product decision. A practical implementation would give the application a least-privilege IAM role, place private workloads behind controlled network egress, and keep administrative access separate from customer-data access. A VPC is a network boundary; it is not, by itself, proof that an application has no public path.
For S3 objects that need customer-controlled key policy and auditability, SSE-KMS is a considered choice. AWS requires the KMS key and bucket to be in the same Region, and a customer-managed key adds control over key policy and grants. That control also creates work: key lifecycle, rotation, access review, recovery, and deletion must be owned. Encryption reduces exposure; it does not decide who may use decrypted data.
Make rights and incidents operational
Access and correction requests expose whether the architecture is real. For an Australian path, that means designing for APP 12 access and APP 13 correction. The application needs a subject-to-record index that can locate primary records, S3 objects, logs, backups, and replicas without handing a caller raw internal identifiers. Deletion or de-identification should follow the retention policy across those copies, while preserving only the records that a documented obligation permits or requires.
CloudTrail can record AWS API activity, which is useful evidence for an investigation. It is not a complete application audit trail: the system still needs events for user access, purpose, administrative actions, and the outcome of rights requests. If a breach is suspected, the client’s incident process and counsel must decide whether the Australian Notifiable Data Breaches scheme applies and what response is required. The architecture supports that review; it cannot make the legal determination.
What this case supports
These controls need verification against the chosen services and the client’s data flow. No new production measurements or legal acceptance are reported here; monthly volume, cost, latency, uptime, production testing, service-by-service regional behavior, retention configuration, rights workflows, and incident handling remain deployment-specific checks.
For the broader architecture and privacy-control discussion, read the whitepaper. Technical controls can be designed and checked, while legal interpretation and sign-off remain with the client’s counsel.