The seam where the model goes

DB.ai is a new product under my Lunvex Labs studio. You describe your shop in plain English — what you sell, what you keep losing track of — and it builds a working inventory system: a schema, seeded stock and supplier tables, low-stock alerts, and barcode product and shipping labels you can actually print. The back office, generated.

I shipped the 0→1 demo this week. There is no AI model in it yet, and that was the point.

One function, swappable

The whole product depends on a single function: generateSystem(input) → GeneratedSystem. Everything downstream — the tables, the low-stock badges, the label studio — only knows about the shape it returns. It has no idea where that shape comes from.

Right now it comes from curated examples. The input keyword-routes to one of three hand-authored systems — a candle shop, an auto-parts store, a vintage reseller. Later, a real model drops into that same function and returns the same shape. Nothing else in the codebase changes. The seam is the entire architecture.

Why I didn't wire a model yet

Because the interesting risk in a 0→1 demo is not the model. It is whether the thing the model produces is worth producing at all.

So I built the back office first. If the arc — describe a business, watch a schema appear, edit a real table, print a real label — is not compelling with data I wrote by hand, no model would save it. Prove the product is worth generating before you spend anything generating it. The model is the part I am least worried about.

Honesty in the demo

The generation sequence is theatre — keywords surface, reasoning streams, tables assemble. But it is labelled theatre. There is no fake "AI is thinking" spinner pretending to call an API that isn't there. The examples are curated and the seam is real and empty. Same discipline as the quant work: don't dress a stand-in up as the real thing.

A note on printing

Labels print without a PDF pipeline. A body.db-print-mode class flips CSS visibility so only the label area survives window.print(), then cleans up on afterprint. No library, no server round-trip — the browser already knows how to print a sheet of paper. Most of the hard parts of "make it printable" are just deciding what not to print.

The 0→1 demo is live — describe a business and watch the system assemble, at lunvexlabs.com/db.

The model is the easy part. The seam is the product.