Q&A

A portfolio build where you watch software get made: pick or describe a small business app, and the site plans it, builds it, tests it and ships it in front of you — ending in a working app you can use immediately and share by URL. It's the "watch the factory work" piece of a three-part showcase for uang.ai ("The AI Factory").

Because "bespoke software built with AI" is easy to claim and hard to show. UangLite shows an agent operating a real ERP; UangFactory (the website) proves instead of tells; UangForge goes all the way: it demonstrates the delivery loop itself — the thing uang.ai literally names itself after — live, in the browser, in under a minute.

No. Independent portfolio work by Rico — not affiliated with or endorsed by uang.ai.

Here is the honest breakdown: the LLM is the planner — it fills a strict JSON blueprint (entities, fields, views, metrics, workflow) from your description. A deterministic runtime — ordinary, tested React code — builds the actual app from that blueprint. So yes, the app you see is really built and really runs; no, the model doesn't write arbitrary code. That division is the whole point.

Because then the demo sometimes works — and a factory that sometimes works is a demo, not a factory. Constraining the model to a schema inside a fixed runtime means every build succeeds, every build is testable, and nothing can be faked. It's also how a lot of real production AI works: model the judgment, engineering the rest.

The runtime's full vocabulary: up to 2 entities with 2–7 typed fields, table/board/calendar views, count/sum/avg dashboard cards, seed data, a create-record form, and one approval workflow with a human confirm. Membership trackers, leave approvals, bug triage, inventory reorders — that class of app. The factory refuses what's outside the vocabulary instead of pretending.

Yes — it's the part most similar demos fake. The checks execute actual validations against your blueprint: schema bounds, every field typed, views resolvable against entities, aggregations resolvable, workflow statuses distinct and present. A failing check stops the line; the fallback is a labeled curated template, never a fabricated result.

You get the nearest match from the curated template library, clearly labeled "curated template". With a key, the planner plans from any description and the result is labeled "planned by LLM". If the planner fails validation twice, you get the curated fallback plus a note explaining exactly what happened.

In your browser. Records you create persist in localStorage under a key derived from the app's blueprint. There is no database. The whole built app is encoded inside its URL, so a link can never leak data you entered after opening it.

Only your prompt (what you typed). No records you create in a built app are ever sent anywhere — they never leave your browser.

Next.js 16 (App Router) · React 19 · Tailwind CSS 4 · TypeScript · AI SDK with any OpenAI-compatible endpoint · zod for the blueprint schema · a URL codec (deflate + base64url) so apps live in links · Vitest + Playwright for tests. No database by design.

The generation strategy was spiked first: 8/8 prompts produced schema-valid blueprints on the first try with a small model. Unit tests cover the schema (including workflow sanity rules), the URL codec round-trip, and seed determinism. Playwright end-to-end tests build an app through the real theater, create a record through the real form, run the approval flow, and open a shared URL.

Yes — open source. Set LLM_BASE_URL / LLM_API_KEY / LLM_MODEL to any OpenAI-compatible endpoint (including a local one). Without them the curated library still powers the whole experience.