Don't read about the AI Factory. Run it.
Describe a small business app and watch it get planned, built, tested and shipped in front of you — then actually use it. The LLM fills a constrained blueprint; a deterministic runtime builds the app. That is how you build software with AI and keep it honest.
The planner fills a strict blueprint schema — never code. Without an LLM key the nearest curated template is served instead, labeled as such.
…or watch a scenario build
Why this is honest
“Watch AI build an app” demos usually cheat — a video, a simulation, or code that only sometimes compiles. UangForge takes the boring-but-real route:
Plan — the blueprint, not code
The planner LLM fills a strict JSON schema inside a fixed constraint space: entity types, field types, views, metrics, one approval workflow. It never writes code — arbitrary output is impossible by construction.
Build — a deterministic runtime
The same machine interprets every blueprint: schema, seed data, dashboard cards, table/board/calendar views, a create form, an approval gate. Verified modules, no surprises.
Test — real checks, real QC
The test stage executes actual validations against the blueprint: schema bounds, typed fields, renderable views, resolvable aggregations, a sane workflow. A failing check stops the line.
Ship — a working app + a URL
The finished app runs live on the page, persists to your browser, and lives in a shareable URL — the whole blueprint is compressed into the link. No database, nothing stored.
The constraint space
The runtime can build: up to 2 entities with 2–7 typed fields (text, number, currency, date, select, boolean), table/board/calendar views, count/sum/avg dashboard cards, and one approval workflow with a human confirm. That's the whole vocabulary — which is exactly why every build works. Inside the limits, the results are genuinely useful trackers; outside them, the factory says no instead of pretending.