What I Built for a Downtown LA Jewelry Auction House (and What It Taught Me About Automation)

In the middle of October 2026 I got a call from Henrik, who owns Jewelry District Auctions in downtown Los Angeles. He runs live jewelry auctions on eBay Live and TikTok Shop, and he is one of the top jewelry sellers on eBay Live. The auctions were not the problem. Everything around them was. Nine days after his deposit I was at his office training his team on the system I built, and this post is what it does, what I got wrong on the way, and what I think any business owner should take from it.

The Problem Was the Day After the Show

A live auction makes its money in a couple of hours and then hands the team a day of cleanup. Hundreds of buyer emails about shipping, payment and sizing, answered one at a time by whoever was free. Winners who bought a loose stone and wanted it set into a ring or a pendant, which meant a chain of messages between the buyer, the team and the jeweler. Hours of footage from every show with the best moments buried somewhere inside, and nobody with time to go find them. And on the money side, eBay pays out in batches, so the bookkeeper was matching payouts back to sales by hand and had no quick way to see when a payout had been returned.

None of that is a hard problem on its own. It is all repetition, and repetition is what software is for.

What Runs on Its Own Now

Everything lives in one hub the team logs into. Here is what it does without anyone asking.

• The inbox drafts the replies. Every few minutes it reads new buyer mail, checks it against the shop's written policies and writes a reply as a draft. It also labels internal mail and automated notices so they stay out of the way.
• A loose stone becomes a custom piece through a link. When a buyer wins a stone, the team sends them a link. The buyer picks a setting, the system generates three AI design concepts for that exact stone in a few minutes, and the order goes to the jeweler with everything attached.
• Every show cuts its own recap clips. A scheduled job pulls the eBay Live replay, finds the biggest sales and cuts the last 30 seconds of bidding plus eBay's own sold screen, biggest sale first. The team wakes up to clips ready to post.
• A daily report lands every morning. Nobody has to go pull the numbers.
• Inventory imports from the sheet the team already keeps, with alerts when something needs a look.
• The books match themselves. Through the eBay Finances API, every sale is matched to its payout and pushed to a sheet laid out the way the bookkeeper already works. A payout that gets returned is flagged instead of found later.
• The team has a shared ChatGPT workspace for everything that does not fit a button.

Two more pieces, eBay sales messaging and Instagram DM replies, are built and waiting on platform approvals. The stack is plain: Vercel and Supabase for the hub, OpenAI for the drafting and the design concepts, and the eBay Finances API for the money. The full breakdown with the timeline is in the case study.

What Still Needs a Human

The inbox runs in draft mode on purpose. It writes the reply, a person reads it, fixes anything that is off and clicks send. Nothing a buyer reads goes out without someone on the team seeing it first. The custom jewelry flow works the same way. The system makes the concepts, but the jeweler and the buyer decide what gets made. The clips get cut automatically, and a person picks which ones to post and where.

I build it this way every time, and I would build it this way for you. A reply with the shop's name on it is a promise. The draft gets the team most of the way there, and the click keeps them honest. It costs a few seconds per email, and it means nobody ever has to explain a message the software sent on its own.

Two Things I Got Wrong, and the Fixes

The first live run of the inbox drafted replies to shipping notices and to the team's own internal mail. It was doing exactly what I told it to do, which was reply to new mail, and I had not told it which mail is not a conversation. The fix is a rule that now runs before anything else: mail from the team, mail from the shop's own addresses and automated notices from carriers and platforms get a label and never get a draft. Simple, and it should have been there on day one.

The second was the clips. eBay logs a time for every sale, and the first version of the clip tool cut around that time. Then I watched a clip and the hammer fell more than a minute after the log said it did. The logged time can be off by minutes, so the tool now watches the replay itself for eBay's sold screen and cuts from there. If it cannot find the sold screen for a lot, it does not fake one. It moves to the next biggest sale and tries the missing one again later.

Both mistakes had the same shape. I trusted a source that was almost right, and almost right is the expensive kind of wrong when the output goes to a customer or onto a public feed. Now I assume every feed lies a little and I check the thing itself.

What a Business Owner Should Take From This

Automate the parts that repeat. Not the whole business, the parts that look the same every day. Replies to the same questions, clips from the same kind of footage, payouts matched to sales. If you can write down how you do it, software can do the first draft.

Keep approval on anything that touches a customer. Let the system write and let a person send. You lose a few seconds per message and you keep your voice, your judgment and your name. The same rule holds for what gets posted and what gets made.

Measure by hours saved. The question is not how clever the system is. It is how many hours the team got back this week and what they did with them. For a live auction business that is more time on the shows, which is the part that actually makes money.

If your business has a day after the show, whether that is emails, footage, orders or books, this is the kind of system I build. Book a call or text me at (805) 433 4465 and tell me what your team does by hand every day. That list is usually the whole plan.