Luna

This one's personal. My wife is a maker — she runs a small hand-crafted jewelry business, and the happiest version of her day is spent designing and assembling pieces, not reconciling stock or generating invoices. But a growing business needs both, and the business side was steadily eating the time she'd rather spend creating.

So I set out to build her the system I'd want running a real operation — and then to put an AI in front of it that could shoulder the parts she dreads, so the toil mostly disappears and she gets her craft back.

Under the hood, Wonder Objects is a genuine commerce platform. It tracks finished products and the individual components they're assembled from, with accurate inventory at every stage of the lifecycle — from the suppliers she buys raw materials from, through the many variations she designs, to every channel she sells through, from her online storefront to Etsy, local shops, and trade shows.

It handles purchase orders with prorated landed cost, price lots, customers, vendors, and sales invoices — the unglamorous machinery a real business runs on, finally accurate and in one place instead of scattered across a fragile spreadsheet.

Luna lives on every page as a slide-out assistant that always knows what my wife is looking at. She answers questions across the whole business — what's in stock, what sold, which customer bought what — and she takes on the creative chores my wife dreads most: naming new pieces, and drafting descriptions, captions, and social posts directly from the product photos.

Because she can see the real sales history, Luna also understands the people behind it — what someone has bought, when and where, and what they (or the people in their lives) might love next. She's a quiet, capable partner for the entire business side of being a maker.

Even for a jewelry shop, I built Luna the way I build trading systems. Reads are free, but there's exactly one write tool — field-whitelisted, returning an old/new diff that doubles as an audit trail, and every change requires my wife's in-chat approval. Luna can suggest updating a price; she can't change one on her own. Architecturally it's deliberately spare: targeted database tools instead of RAG, on the OpenAI Responses API, at about half a cent per conversation.

That discipline paid off in an unexpected way — Luna became an accidental data-quality auditor. She flagged an inflated invoice count that exposed a counting bug, and surfaced corrupt records like a sale dated to the year 0112, making the inventory more trustworthy, not less.

The measure of success here isn't a benchmark — it's that my wife stopped dreading the business side of her craft. Luna handles the parts she never wanted to do, the inventory is finally trustworthy, and she's back to spending her time where it belongs: designing and making beautiful things.

It's the most rewarding thing I've built — proof that the same care I bring to high-stakes infrastructure can make one person's working life genuinely better.