Goldsmith: AI design plus a marketplace for custom jewelry

Custom jewelry is expensive mostly because of the people between you and the person who makes it.   Jewelers hold the relationships with manufacturers and designers, and they mark prices up to cover that, plus the store and the hours of consulting that custom work takes.   A buyer can’t tell whether a price is fair, so many don’t trust the process, and custom pieces stay out of reach for anyone who isn’t already wealthy.   I described Goldsmith as the Warby Parker of custom jewelry.

I wrote the product plan, the research plan and the business model, built the financial model, presented the plan to investors in December 2024, and delivered the platform in 12 weeks.

Three things about the market stood out.   The first is price: I found that custom work is marked up by 200% to 500%, and the buyer sees only the final number.   The second is trust: most people don’t buy custom jewelry often enough to know what a fair price looks like, so they either overpay or give up.   The third is access: a small jeweler without a strong network of verified manufacturers, or without expensive hardware for 3D design, can’t compete on speed or choice.

Nobody agrees on the size of the custom slice.   Published estimates of all jewelry sales in 2024 sit at about $367 billion to $371 billion.   Estimates for custom jewelry services alone run from about $1.09 billion to $3.63 billion depending on the publisher, and broader definitions of customized jewelry reach $31.8 billion to $36.2 billion.   The gap comes from what each report counts, so I treated the range as a reason to size the opportunity from the bottom up, using manufacturers and buyers I could actually reach, and not from a market report.

The business has three sides, and each one has a different reason to use it.   I wrote them down early because they decide what the product has to do.

I wrote a customer discovery interview guide.   It starts from a stated problem and value hypothesis, then asks buyers ten questions: how often they buy, what hurts most about custom jewelry, how much they trust a jeweler’s price on a 1 to 10 scale, whether they would design with natural language, whether they would choose a piece from a 3D preview, and how long they would wait.   I built it so the answers would turn into simple probability models and descriptive statistics, not a pile of quotes.

I don’t have interview results to report, and I won’t imply that I do.   The guide is the work product.

I looked at the tools jewelers use to design.   They are all CAD software, and the closest existing route to a finished piece belongs to one company: Gemvision’s MatrixGold is owned by Stuller, which also makes custom jewelry from submitted designs, but only for businesses with a resale account.

In my 2024 review none of these paired design from a text description with a marketplace where manufacturers compete on the same file.   That is the gap I built Goldsmith around.   It is a claim from my review, not a survey of every tool on the market.

Goldsmith has two parts, a design tool and a marketplace, and they depend on each other.   The tool creates the design a manufacturer can quote on.   The marketplace turns that design into a real piece at a price the buyer can compare.

Most of the value in a design tool is the small facts a buyer or jeweler would otherwise look up elsewhere.   The tool keeps three reference tables one click away, so the design, the stone and the size always agree.   These are standard industry reference values, not figures from the plan.

A design check reads the same numbers.   The rules below are example starting values.   Each manufacturer’s own limits replace them.

The marketplace only works if manufacturers can make what the tool designs.   I checked who actually offers the services the platform would route to, and every company below publishes what it does.   None of them is a partner, and none has quoted anything.

This is the platform I built in 12 weeks.   Buyers and small and mid-size jewelers are on one side and verified manufacturers, the supply side it routes to, are on the other, with the design engine and reference data in between.

A buyer describes a piece in plain language.   The platform creates candidate 3D designs with a language model and generative CAD on open source design software.   The buyer refines them, or a jeweler with the full tool edits by hand on cloud GPUs.   Reference tables for ring size, carat to millimetre and metal purity, and manufacturing rule checks, keep the design realistic.   The rules are example starting values that each manufacturer’s own limits replace.

The design then goes to verified manufacturers, who share business documents before they can bid and show certifications where held.   They quote through a request for quotation with estimated cost, the final design and specifications.   The buyer compares the quotes, receives 3D printed models of the top candidates to try, picks one, and expects the finished piece in about three weeks.   That timing is the plan’s expectation, not a measured result.

Two loops matter.   Design checks send fixes back into the model, and quotes and try-on results send the buyer back to refine the design.   Underneath sit payments, and training the models on data that is not someone else’s design, with a way for a buyer to protect a design they create.

I kept the build to 12 weeks by deciding early what not to build.   I used open source design software instead of writing a modeler, and put the AI on top of it.   The full tool runs on cloud GPUs, so no jeweler needs special hardware.   The reference tables and manufacturing rules are data the design is checked against, and manufacturers replace the example limits with their own.   The design file is the one thing every manufacturer quotes on, so there is no custom work per quote.

The plan had five revenue streams, each tied to a different customer.   They were not meant to start at once.

Buyers come in through the website and social media.   Jewelers and manufacturers are reached through LinkedIn, trade expos, follow up calls and account managers, with HubSpot connecting the channels.   I expected social media outreach to be the cheapest per contact and account managers to matter most for jewelers and manufacturers, since those are relationships, not clicks.

I wouldn’t release the whole platform at once.   A first release needs to answer one question: will a buyer pick a final design from a 3D preview and wait about three weeks for it?   So I would limit it to a small run.

Goldsmith went from a concept to a business model, a financial model, an investor pitch and a working platform in 12 weeks.   It has not been released to the public, so I have no usage or revenue to report, and every figure in the plan below is a target, not an outcome.   I would rather say that plainly than dress up targets as results.

At the December 2024 pitch I presented a plan for a seed round.   These are targets from that plan, not results.

The plan rests on one behavior.   Buyers will choose a final design from a 3D preview and wait about three weeks for the finished piece.   If they won’t, the cost advantage disappears, because a jeweler’s fitting room is the thing I was trying to replace.   So before a public release I would run the small version above with a few manufacturers and a few dozen real buyers.

Before a run like that I would write down what counts as a pass.   The criteria below are proposals I would set in advance, not findings.

My business model work ended with two priorities to validate, and I still think they were right.