Intent. Intelligence. Impact. / Creator Lab

Day 2. Wicked Spooky's architecture review

Plan → Execute → Certify → Impact. Wicked Spooky reviews the AWS features against Crisis Cat's budget, integration assumptions and ambition to reach thousands of subscribers.

The architect's commitment

“Crisis Cat, I am responsible for the cloud environment that will make your design deliverable. My 30 years in IT help me recognize the operational problems; SAA-C03 study gives us the AWS vocabulary and architectural choices. We will prove those choices with evidence. We will design for the viral moment before it happens, without paying for speculative capacity today.”

We applied iLab's product design journey, not a service shopping list. Discover separated the pitch from facts. Plan selected boundaries, dependencies and success criteria. Execute means bounded implementation experiments. Certify means measured acceptance. Impact checks whether the product earns enough to operate and produces useful outcomes. A failure returns us to discovery and planning with better evidence.

Crisis Cat's “everything is in place” assumption

The scenario assumes necessary infrastructure and integrations exist. The architect does not treat that assumption as deployment evidence. We turned it into this readiness ledger. Every row is unverified in this design lab; live delivery depends on the actual proof.

All supported platforms and billing capabilities must be ready before that capability is sold, rather than rushed into existence after signups arrive. Missing access narrows the advertised supported set and becomes a recorded design event. No backend resource, provider account, credential, checkout or commercial commitment is created by reading this lab.

AWS feature review

From almost no budget to paid growth

We chose one paid AWS account, one primary Region (us-east-1), static presentation, queues and managed services with no speculative always-running fleet. Training uses synthetic data, dry-run publishing, local work where useful and small live experiments. A proposed $10/month AWS cash target is a planning envelope, not a guarantee; the founder needs runway for secrets, storage, logs and other charges even before paying customers.

AWS's post-July-15-2025 Free Tier separates temporary credits from recurring service allowances. New eligible accounts receive $100 and can earn up to $100 more; credit validity is 12 months from account creation. The Free account plan's maximum six-month period is a different rule. This owner already chose the Paid plan, so usage beyond applicable offers/credits is billable. Joining/creating an AWS Organization can invalidate credits. Verify the account rather than infer benefits from older service-page examples. AWS Free Tier FAQ

Paid subscriptions are priced against full replacement cost, not a temporary discounted bill. The operations case study contains the complete assumptions, arithmetic and subscription hypotheses. Save a dated Pricing Calculator estimate before each experiment and refresh the rate card before implementing. Model availability and fees can change.

Local AI: what we selected and what we rejected

The inference contract records model/version, task, tenant, source IDs, token or compute usage, output schema and evaluation result. A task deadline determines whether an unavailable local worker is requeued, escalated within the reserved budget, or returned for review. Lease expiry permits recovery without two workers committing the same result. Infrastructure roles enforce authority; model outputs cannot grant it.

Options and architecture decisions

What happens when the marketing catches fire?

We did not design a magical free-to-paid conversion at the moment of virality. On the owner's Paid account, applicable metered usage simply becomes billable after offers. Resources scale inside configured bounds. The business admits paid, budget-reserved work; queues absorb bursts; caps, downstream quotas and platform limits determine completion time.

flowchart TD
    accTitle: Funded growth with bounded autoscaling
    accDescr: Signups create server-verified entitlements. Requests reserve cost before durable queue admission. Workers scale within AWS, model and platform limits and report outcomes and margin.
    A[New subscribers and viral traffic] --> B[Verify identity and paid entitlement]
    B --> C{Reserve usage and worst-case cost?}
    C -->|No| D[Offer upgrade, economical option or retry]
    C -->|Yes| E[Durable queue admission and job ID]
    E --> F[Workers scale within configured paid envelope]
    F --> G[Respect model quotas and platform rate buckets]
    G --> H[Approval check and idempotent state transition]
    H --> I[Completion, receipt or actionable failure]
    I --> J[Settle usage and reconcile billing]
    J --> K{Margin and backlog healthy?}
    K -->|Yes| F
    K -->|No| L[Bound admission, investigate and adjust]

flowchart TD
    accTitle: Funded growth with bounded autoscaling
    accDescr: Signups create server-verified entitlements. Requests reserve cost before durable queue admission. Workers scale within AWS, model and platform limits and report outcomes and margin.
    A[New subscribers and viral traffic] --> B[Verify identity and paid entitlement]
    B --> C{Reserve usage and worst-case cost?}
    C -->|No| D[Offer upgrade, economical option or retry]
    C -->|Yes| E[Durable queue admission and job ID]
    E --> F[Workers scale within configured paid envelope]
    F --> G[Respect model quotas and platform rate buckets]
    G --> H[Approval check and idempotent state transition]
    H --> I[Completion, receipt or actionable failure]
    I --> J[Settle usage and reconcile billing]
    J --> K{Margin and backlog healthy?}
    K -->|Yes| F
    K -->|No| L[Bound admission, investigate and adjust]

Training generation concurrency begins at 2; a commercial experiment begins at 10 with a per-tenant outstanding-job ceiling of 20. These are proposed settings, not AWS default quotas or tested capacity. Reserve separate capacity for approvals, verified billing/webhooks and reconciliation so generation cannot starve trust-critical paths.

Effective throughput is the minimum of worker capacity, database access-pattern capacity, model request/token quota, provider rate allowance and funded usage. A concurrency cap alone does not enforce requests/minute. Warmup, backlog age, throttles, deadlines and drain time are measured with one-tenant, 100-tenant and 10× burst scenarios before promising thousands of customers. Actual new-account quotas may differ from published defaults. Lambda scaling

DynamoDB on-demand maximum throughput targets offer useful protection but are best effort and can be exceeded temporarily by burst capacity; they are not a hard dollar cap. AWS throughput guidance Immediate commercial controls remain atomic reservations, bounded tokens/bytes/retries, worker limits and circuit breakers. Quota requests and architecture changes are planned before a launch, not assumed instantaneous during a spike.

Subscription economics and the profitability guardrail

We retained the 30% runtime gross-margin floor and selected a 50% operating contribution-margin target after allocated support, fixed cost and a maintenance reserve. A 30% markup yields only 23.08% gross margin. The floor requires net revenue at least runtime cost / 0.70; the operating target requires roughly a 100% markup on all-in cost before payment fees and refund allowance.

All-in cost = fixed monthly cost / conservative paid subscriber floor
              + included-use cost envelope + support + reserve
Minimum price = (all-in cost / 0.50 + fixed payment fee)
                / (1 - variable fee rate - refund reserve rate)

Prior hypotheses remain $19 Solo (one brand, two channels, 120 text drafts, 60 destination publications, 30 briefs, 1 GB media) and $49 Studio (three brands, six connections, 400 drafts, 240 destination publications, 90 briefs, 5 GB media). They are not offers on this lab site. With assumed 3% variable fees, $0.30 fixed fees and a 2% refund reserve, net revenue is $17.75/$46.25. All-in envelopes of $8/$21 yield about 54.9%/54.6% contribution margin; measured costs and actual fee terms must validate those envelopes.

No speculative viral subscriber count reduces today's fixed-cost allocation. If few subscribers cannot cover the minimum cost, the founder funds runway, changes pricing or delays launch. Heavy rendering, premium models and media egress are explicit add-ons. Overage charging requires opt-in; advertised included use is enforced server-side.

Each admitted job reserves worst-case cost and usage atomically. At 70% allowance we notify, at 90% we offer upgrade/overage options, and at exhaustion we reject new expensive work or offer a bounded lower-cost path. Forecast runtime margin below 35% triggers investigation; below 30% blocks new subsidized admission. The 50% operating target has its own stricter envelope. Accepted work keeps its reservation and visible status. Billing lag and ambiguous third-party charges require contingency; AWS budget alerts are not a hard stop.

Observability, resiliency and recovery

CloudWatch observes backlog age/depth, DLQs, function errors/throttles/duration, workflow failures, model tokens/latency, hot keys, unknown publication receipts, due publication delay and cost/job. Approval, platform and billing events carry redacted correlation IDs. Durable minimal business audit is separate from short-retention debug logs; secrets and private messages are not copied into evidence.

Design targets are p95 acknowledgement <2 seconds at declared load, 99% of due publications attempted within five minutes when providers are available, and briefs ready by 08:00 tenant local time. They are unmeasured. Provider outage, exhausted entitlement and unknown delivery remain visible. Proposed business-data RPO is <=24 hours for synthetic training and <=15 minutes for commercial service, with regional restore RTO <=4 hours. A Region-loss commitment needs cross-Region protection and a timed restore; regional PITR alone is insufficient.

Retries are bounded and jittered. Permanent authentication/permission errors stop; timeout after an external send is reconciled first. DLQ redrive rechecks current approval, entitlement and provider receipts. Duplicates are expected across event boundaries; conditional state and deterministic keys handle them. We do not promise exactly-once publication across independent systems.

Quality inputs, replaceable models

We designed the workbench around a portable input package: objective, source evidence and dates, brand profile, permissions, constraints, expected schema, evaluation rubric and cost allowance. A provider adapter translates that package into the chosen model API and normalizes the output. Claude, a licensed local model, or another supported cloud model can be evaluated against the same acceptance criteria. We select models by measured quality, privacy, tool support, latency and total runtime cost. Model agnostic does not mean every model behaves identically; compatibility and regression evidence determine which adapters are enabled.

Vendor links help readers explore the tools. Partner referrals and implementation services are supplementary business opportunities; the subscription guardrail assumes no referral income. Anthropic's Claude Partner Network provides one documented partner route for delivery expertise and registered business referrals.

All 14 SAA-C03 task statements

The official guide weights secure/resilient/high-performing/cost-optimized domains at 30%/26%/24%/20%. This case study maps each task to its design application and evidence requirements.

The detailed per-skill acceptance register and extension plans remain in the operations case study. Extension work is bounded and torn down after testing. Account-organization exercises are isolated/deferred when necessary to preserve credit eligibility. Designing unnecessary services into Creator would increase cost; studying and demonstrating those options is still required for broad exam coverage.

Fail-forward: evidence changes the design

No incident is fabricated. Each real failure records requirement, observed symptom, impact, detection signal, competing hypotheses, root-cause evidence, containment, alternatives, fix, rollback, verification and the affected SAA skills. This is Continuous Learning applied to a concrete delivery decision.

The event is only “verified” after the original failure case and relevant regressions pass with linked artifacts. Symptoms, unsuccessful hypotheses and cost impact stay in the record. Fail-forward means controlled repair and demonstrated learning, not proceeding with unresolved unauthorized publication or unbounded spend.

Day 2 handoff: ready to build the right thing

We closed the design review with the AWS feature matrix, integration readiness ledger, budget/paid-growth distinction, local AI contract, explicit alternatives, subscription guardrails and exam evidence plan. The next implementation slice begins with identity, tenant state, observability and dry-run research/content; subsequent slices prove approval, publishing, subscriptions, attribution and recovery before commercial enablement.

The lab will evolve as those experiments generate evidence. The eventual Creator Workbench is built from these specifications; this training section remains its discussion, decision and learning record.

Lesson Markdown · Wicked Spooky IT / delivery identity