From requirements to architecture — and the GenAI Lens
Lesson 1 of 4 in Designing GenAI Solutions and Selecting Models.
The professional move the whole exam keeps re-testing is turning business requirements into an architecture — which FM, which integration pattern, which deployment strategy — and proving it cheaply before committing. Requirements come in two kinds. Hard constraints (data residency, latency SLOs, compliance regimes, freshness) are filters: an architecture that violates one is wrong no matter how elegant. Preferences (cost, effort, time-to-market) are tiebreakers the qualifier arbitrates.
The proving ground is a proof of concept on Bedrock: serverless, pay-per-token access to many models means feasibility, performance characteristics, and business value get validated in days — prompt in the console playground, then the same call through the API — before anyone provisions infrastructure or signs a model strategy. The AIP-C01 exam guide names exactly this: validate before full-scale deployment.
Requirements to architecture, with the PoC gate in the middle
- Business need
What outcome, for whom, worth how much — scoping is phase one of the GenAI lifecycle in the Well-Architected lens.
- Frame requirements
Separate hard constraints (residency, latency, compliance, freshness) from preferences (cost, effort). Constraints filter; preferences rank.
- PoC on Bedrock
Serverless and pay-per-token: test candidate models against real task samples. Measure quality, latency, and token economics — not vibes.
- Proves value?
The cheapest moment to kill a GenAI project is before the architecture exists.
- Re-scope or stop
A failed PoC that cost a week of prompting is a success of the method.
- Harden the design
Integration pattern, deployment strategy, safety and observability — the PoC becomes an architecture with standardized, reusable components.
- GenAI Lens review
Review the design against the Well-Architected Generative AI Lens — six pillars across the GenAI lifecycle, runnable as a lens in the WA Tool.
What counts as a "standardized technical component"?
Reusable building blocks that make the third and fourth deployments look like the first: a shared model-gateway layer (one place for auth, logging, model switching), approved prompt templates in Prompt Management, IaC modules for knowledge bases and guardrails, and a common evaluation harness. The exam rewards reuse over reinvention whenever consistency is a stated need.
Why PoC on Bedrock even if production might land elsewhere?
Zero infrastructure and per-token billing make it the cheapest feasibility instrument on AWS: compare several FMs against the same task sample in hours. If production later needs self-hosted serving, the PoC still answered the expensive questions — can a model do the task, at what quality, at what token cost.
What does a professional PoC measure?
Three numbers, not one demo: quality on a representative task sample (rubric or LLM-as-a-judge), latency at realistic payload sizes, and token economics per request. Those three carry straight into model selection in the next lesson.
Interactive checkpoint quiz (2 questions) — open this page in a browser to take it.