AWS

BtSavvy's posture on AWS

AWS is the most mature cloud for production AI systems by some distance. The provider's depth across managed model services, compute, storage, networking, and observability gives engineering teams the broadest set of working primitives to build against, and the deepest set of failure modes to design around. That breadth is an asset and a discipline cost simultaneously: the same surface area that makes AWS a complete substrate also makes provider-specific accidental complexity easy to accumulate.

BtSavvy's posture on AWS is vendor-aware on implementation, vendor-neutral on outcome. FORGE runs on AWS where AWS is the right substrate, and on Azure or GCP where those are. The platform is identical across providers; the implementation choices differ. This page describes the working posture when AWS is the chosen cloud.

Which services carry AWS

All eight FORGE services apply on AWS. Three carry AWS-specific depth most often:

Cloud. Cloud platform engineering across landing zones, organisation-level guardrails, network architecture, and provider-native observability. AWS Control Tower, Service Control Policies, AWS Config rules, and Security Hub are the primary platform-level primitives; Transit Gateway, PrivateLink, and Resource Access Manager the primary network primitives.

DevSecOps. CI/CD on AWS-native or third-party tooling: CodePipeline, CodeBuild, GitHub Actions on AWS-hosted runners, GitLab CI with AWS integrations. Pipeline-vending discipline applies regardless of tool; AWS-specific work concentrates in the IAM model, artefact-storage discipline, and production-evidence logging through CloudTrail and EventBridge.

Build. Application engineering on AWS-native runtimes: EKS for containers, Lambda for event-driven systems, Bedrock for managed model invocation, SageMaker for full-lifecycle ML. The firm builds against AWS's primitives directly rather than through a layer of provider-abstraction, which is a debt that compounds against the engagement.

FORGE on AWS

FORGE is provider-neutral as a platform. What changes between providers is the implementation surface; what stays constant is the discipline.

Factory on AWS produces the customised builder against AWS's APIs and an accelerator library calibrated against AWS primitives: IAM policy authoring, VPC design patterns, Bedrock guardrail configuration, SageMaker pipeline construction.

Observability on AWS uses CloudWatch and X-Ray as the substrate. Retrieval-quality, guardrail-trigger, and calibration-skew signals from AI workloads are instrumented as structured CloudWatch metrics from the first build: first-class SRE signals on the same plane as latency and error rate.

Review on AWS produces signed gate logs against AWS-specific gates: landing-zone provisioning, IAM policy promotion, Bedrock model access grants, SageMaker model registration, production deployment. Each gate carries the accountable engineer's name in the queryable record.

Governance and Trust on AWS uses AWS Config, Security Hub, and Audit Manager as the policy-as-code substrate, with EU AI Act and ISO/IEC 42001 conformance evidence emitted continuously rather than assembled retrospectively for an audit.

Expert Intelligence on AWS uses Bedrock for frontier-model access, SageMaker for fine-tuning and custom-model serving, and OpenSearch and Bedrock Knowledge Bases for retrieval. The firm fine-tunes against the model providers AWS surfaces (Anthropic, Meta, Mistral, AI21, AWS's own families) and against externally hosted models where the workload requires it.

Engagement-pattern sketches

Representative patterns FORGE is built to deliver on AWS:

A landing-zone establishment with FORGE governance from day one. AWS Control Tower plus organisation-level Service Control Policies, AWS Config conformance packs, and Security Hub as the policy-as-code spine; the first FORGE workload deployed onto the established landing zone with full observability and gate-log discipline from the first commit.

A Bedrock-based AI system that emits EU AI Act conformance evidence continuously. A frontier-model deployment using Bedrock as the substrate, with retrieval-quality and guardrail-trigger telemetry as live SRE signals, the EU AI Act Article 14 human-oversight controls enforced as a promotion gate via policy-as-code (AWS Config checks the required oversight controls are present and active), and the full traceability genome (the requirements-to-prompt-to-deployment chain carried in the traceability matrix and the signed gate log) bound end to end.

A SageMaker-based domain-fine-tuned system with signed model promotion. Targeted fine-tuning of a frontier model into a domain expert, deployed via SageMaker with the accountable engineer recorded against every model promotion in the gate log.