You're building one integrated platform for Nino Kvaratskhelia's family winery in Sighnaghi, Georgia — three connected sub-systems on one system: tourism bookings, wholesale order management across twelve countries of compliance, and direct-to-consumer e-commerce with age verification and shipping restrictions.
What you're practicing: producing the architecture and product spec from a non-technical ambition; designing the boundaries between three sub-systems that share one inventory truth without coupling; designing the verification architecture, the observability and SLO model, the deployment topology, and the AI development infrastructure from scratch; deploying, monitoring on real signal, negotiating a scope line, and writing a handover packet a future developer and a fresh AI session can inherit.
What you're learning about the loop: AI builds the pieces fast and holds none of the system's coherence. It produces a checkout that ignores the allocation model, a wholesale endpoint that skips per-country compliance, infrastructure sized for a demo that fails under the harvest spike. It reaches for Kubernetes for a twenty-person winery, and it builds every "can we also add a wine club?" the moment it's mentioned. You hold the global coherence AI cannot, size the system to a rural-Georgia reality, and know which pieces — the allocation model, the compliance logic, the trust boundaries — not to delegate at all.
Your Role
You are the architect, the build director, the operator, and the only technical person on the engagement. The promise: by the end you can own a system — its architecture, its trade-offs, its scope, its production life, its handover — while directing an AI that builds fast and holds nothing.
This is the thinnest the work ever gets. Every earlier project handed you some structure — a spec, a breakdown, a verification target, someone to escalate to. Here there is none of that: no reviewer, no escalation path, no one above you. You produce every one of those, because the track has spent twenty-three projects moving that structure from the platform into you. This is where you supply it. The single exception is narrow and changes nothing about that ownership: an independent architect is reachable on demand in chat for one question only — whether a small rural winery's system should be distributed at all — and he never makes the call or hands you an answer.
The AI relationship is yours to design before any code is written: the project memory, the path-scoped rules, the skills, the delegation structure. You decide what to implement directly, what to hand to one agent, what to split across several, and what is too load-bearing to delegate. AI is the safety rope on the hardest terrain in the track — you build the anchor points.
What's New
Last time you ran an evaluation, not a build: you assessed what transfers between AI development tools under a vendor-independence mandate and produced a portable framework, a migration-risk assessment, and a two-audience presentation. There was no system to build — the method was the deliverable, and its shape was named for you.
Now there is no method to hide behind. P24 is a full system, built and deployed and monitored and handed over, and nothing about its shape is given. You run discovery from one email, surface the constraints and blind spots Nino has not thought to mention, and hold the coherence across a multi-week project where a week-one decision still has to be retrievable in week four.
The hard part is honest: a beautiful checkout that ignores the allocation model is not a working system. The deliverable is not any single component — it is the coherence across all of them, proven to hold for someone who arrives after you.
Tools
No new tool category. Everything here you have used before; the difference is that you choose and assemble the stack yourself.
- Claude Code — project memory (CLAUDE.md / AGENTS.md), skills, hooks, and sub-agent delegation, all designed by you up front.
- Git and GitHub — the engagement repo; GitHub Actions runs the verification architecture you design.
- Next.js (App Router, SSR), TypeScript, PostgreSQL with a considered service-boundary model, Stripe for DTC payments.
- Terraform and an AWS/Amplify-class host with a CDN — sized to the actual problem, not AI's default.
- An observability stack (structured logging, metrics, SLO dashboards, alerting), axe-core for accessibility, and a load-testing tool for the seasonal-spike check.
The exact stack is your decision. These are the expected vehicles, not a prescription.
Materials
nino-first-contact.md— Nino's one email: the ambition and the pressures, in her voice. This is the entire brief. It does not spell out the technical answers; discovery does.winery-context.md— background for the engagement: the winery story, qvevri production, the twelve importer countries, the tourism program, vintage volumes (including the 500-bottle Saperavi), the seasonality profile, and Nino's accumulated scope requests. It does not resolve them — those are your calls.compliance-reference.md— reference data you direct AI against: the twelve countries' import constraints, labeling rules, the shipping-restriction matrix, age-verification requirements, and the National Wine Agency certificate-of-origin contract.brand-reference.md— the storefront's visual-identity direction: identity cues, what "feels like us" means to Nino, and the material and texture direction. Direction you translate, not an implementation.
There is no starter code and no scaffold. A reference exemplar CLAUDE.md is included in the materials — read the capstone note at the top and then author your own from the PRD and architecture you produce in Unit 1. You curate it through the rest. By the capstone that is assumed practice, not a handout.