Baby Crafts
Retail · D2CA personalised baby-clothing range turned into an AI engine that puts a child's own name on the product, plus the web-studio built to sell it.
Baby Crafts UK sells handmade, personalised baby and children's clothing — every single order is a one-off, made for one name, one colour, one style choice.
Showing a customer what their own choice will actually look like, for every product and every combination, is the kind of studio work a small brand can't staff or afford to do by hand.
The obvious way to build that is the wrong one. A canvas the customer can drag, scale and rotate on will happily produce a picture the workshop cannot make — an embroidered garment has a production zone, and a name sitting outside it is a preview that becomes an apology.
Generation also costs money on every press. A preview page that rendered as it loaded would spend that money on everyone who was only browsing, which is most people.
And this is a children's brand. Whatever the engine puts on a garment, and whatever a customer uploads into it, has to clear a standard that a general-purpose image tool does not apply by itself.
An AI personalisation engine in three steps: pick how it will be made, pick the exact piece from the real catalogue, then add a name or initials, a thread colour, a lettering style, a little motif and where it sits — and the engine generates a preview of that exact product with that exact personalisation applied, before anything is made.
The decision the whole engine turns on is that the product owns placement, not the customer. Each garment is calibrated once in an internal tool, and the customer chooses the name, the thread, the lettering, the motif — and the position only where that garment genuinely has production alternatives. The engine keeps the geometry to itself: where the lettering sits, how big it is, what angle it takes, how it follows a curve or a fold, and which seams and straps it has to stay clear of. That trades a feature the customer might enjoy for a preview the workshop can actually reproduce.
Products do not all personalise the same way, so the engine routes each one into the mode it belongs to. A garment with a known embroidery zone takes the name into that slot. A template photograph that already carries a sample name has that name replaced in the same place, in the same style, rather than gaining a second name somewhere else. A blank tee takes generated artwork against preset placements. And a product that is personalisable in the real shop but has no calibrated preview yet gets no preview at all — the page says so and points at the real product, because a plausible-looking guess is worse than an honest gap.
Where a garment needs handling of its own, it gets its own rule file, generated into the site rather than hand-edited. The file holds that garment's zones, the options it is allowed to offer, its character limit, and its own instructions to the model — what to do, what to leave untouched, what to keep away from and what would simply be wrong. A smocked babygrow's rules ask for small, low-relief lettering in proportion to a newborn garment, insist the smocking rows and the popper placket survive exactly as they are, and rule out a printed-ink or vinyl finish on something that has to read as stitching.
A full web-studio sits around it — storefront, product catalogue, the personalisation flow, and the back office that keeps it honest: the tools that qualify each garment for the engine, review what it produced and watch how it is behaving.
The catalogue behind it is the client's own range, not a handful of demo items — 323 real products across 15 categories, 214 of them personalisable and 76 already carrying a calibrated blank, so the engine works on the things she actually sells.
Nothing renders when the page loads; the customer presses the button, and MAST built the spending controls around that. Results are cached against the design itself, so an identical set of choices is served back rather than bought again, and the day is capped — 150 renders, of which 40 may go to artwork generation — with a throttle between requests on top. The model key stays on the server as a secret and never reaches the browser.
Because it is a children's brand, the safety work runs three deep: a text gate before any money is spent, the generation model set to block at its lowest harm threshold, and a vision check on both what a customer uploads and what the engine produces before either can be cached or shown. It fails closed. A check that cannot return an answer refuses the image and says so, rather than reading its own silence as approval. A separate readback then compares the finished picture against what the customer actually asked for.
What it actually looks like, running.
One product range now generates its own personalised previews and marketing imagery on demand — the kind of made-to-order studio work a big brand would staff a team for, running for a small one instead.
Because the product owns the placement, what the customer sees is what the workshop can stitch. The preview stops being a picture of a wish and starts being a picture of an order.
The running cost is bounded rather than open-ended: an identical design is served from the cache instead of re-bought, the day has a ceiling, and every generated result is labelled a concept preview rather than a photograph.
Half an hour to see what a personalisation engine looks like for your business.
Start a conversation