❯ git show --details cases/pricing-agent.md
AI Product

Shipping an AI agent where the metric was whether sellers took its advice

An eCommerce analytics and repricing tool for sneaker resellers on StockX and GOAT. I designed and built it on Lovable and Claude. The number I managed to was not GMV. It was acceptance rate: how often a seller took the price the agent proposed.

role
Product lead; built the MVP myself
timeline
2024, three months to GMV
users
US resellers on StockX and GOAT
stack
Lovable, Claude, PostHog, a 200-SKU golden set
outcome
$500K+ GMV; acceptance 42% → 71%

The problem

Resellers hold hundreds of SKUs across two marketplaces where prices move hourly. Pricing by hand means either leaving money on the table or sitting on stock. The existing tools were rules engines that required the seller to know the answer in advance.

The bet was that a language model could read the market the way an experienced seller does, propose a price with a reason, and get better as the seller corrected it. The risk was the opposite: an agent that is confidently wrong on a $400 pair costs a seller real money and they stop listening after the second time.

What I built

01
Golden dataset of 200 SKUsPriced by hand with experienced sellers before the agent proposed anything. Every change to the prompt or orchestration was scored against it.
02
Override as a first-class actionThe review UX made rejecting or editing a price as easy as accepting it, and every override was captured with the seller's price and, where they gave one, the reason.
03
Five instrumented health metricsAcceptance rate, override rate, time to decision, price realised against proposal, and sell-through. Tracked in PostHog from the first week.
04
Weekly eval loopOverrides from the week were added to the golden set, the agent was re-scored, and the orchestration changed only when the score improved.

Results

$500K+seller GMV in three months
42% → 71%acceptance rate, week 1 to week 8
200SKUs in the golden set

Acceptance climbed from 42% in the first week to 71% by week eight. The model did not get better in that time. The evaluation loop and the enablement did. The venture was Founder Institute funded and crossed $500K in seller GMV within three months of launch.

What stayed with me

  • Overrides are signal, not defects. The sellers who overrode most in week one were the ones who trusted the agent most by week eight, because they had taught it.
  • Pick a metric the user feels. GMV was the business number. Acceptance was the number that told me whether the product was working, and it moved first.
  • The eval discipline came first, and that is why the GMV came. I have run every AI product since the same way.
06.

connect()

Building something where a wrong number costs money?
Fintech, accounting, operations, applied AI. I am always open to talking with people who ship.

$ echo "Let's build something that survives an audit"