aws-startups-solution-architecture

This skill should be used when an automated agent's output is published under its own identity and carries weight in someone else's work, such as a reviewer whose verdict lands on a pull request, and the open question is how much authority it has actually earned. Covers which conclusions such an agent may state as settled and which it must hand to a person, why a verdict that moves between runs on unchanged input cannot be allowed to decide anything, handing an uncertain item to a human rather than discarding it, grounding each claim in something computed rather than recalled, and the published-state and recall behaviour that decides whether people keep reading the output or learn to skim it. It should also be used when such an agent is being ignored, contradicts itself between runs, or reports a state that nothing current supports. Not for measuring or improving an agent's own output quality, which belongs to aws-agents:agents-optimize, nor for building, deploying, or hardening an agent.

This skill should be used when designing or fixing tenant isolation for a multi-tenant SaaS on AWS, where a team with no dedicated security engineer must make a cross-tenant leak structurally impossible rather than a code-review responsibility. Covers choosing a silo, pool, or bridge model per layer rather than per application, enforcing isolation in IAM session policies and the database so a forgotten predicate fails closed, partitioning data across DynamoDB, Postgres, and S3, containing noisy neighbors, and attributing cost per tenant in a pooled fleet. It should also be used when hardening an existing pooled deployment against cross-tenant access, or when one large customer demands dedicated infrastructure mid-deal. Not for single-tenant architecture or general IAM policy authoring, which belong to the aws-core skills upstream.

This skill should be used when a startup without a dedicated ML platform team needs recurring, latency-tolerant offline inference for a self-hosted 1B to 8B generative language model on AWS while minimizing GPU spend and recovering automatically from Spot interruption. It prescribes AWS Batch backed by Amazon ECS Managed Instances Spot, with model artifacts staged in Amazon S3, deterministic input shards, vLLM offline inference on one GPU, bounded retries, and output reconciliation before success. Covers sizing from the real prompt and model footprint, choosing networking without an avoidable idle cost floor, and verifying that results are complete and GPU capacity is gone. Not for online endpoints, training, embeddings, other inference engines, multi-GPU or distributed inference, or one-off experiments where manual rerun is acceptable.