Auction OS: a multi-tenant SaaS with web-grounded AI

“Six of us used to run every auction. Now three close it, and the rest of us finally have time to bring in new business.”- the auction house's owner

A complete auction-operations platform built solo in four months: AI cataloguing, multi-marketplace syndication, a full business backoffice, and the go-to-market engine to sell it. Onboarded a real auction house, went live on eBay.com, ~40 features, 79-84% unit margin.

The product is public: auctionos.co.il - the Hebrew marketing site and demo funnel I built to sell it.

Customer details are anonymized; live API keys and account identifiers are kept private.

Problem

Small auction houses run on spreadsheets and list each item by hand, into one marketplace at a time. Cataloguing is the bottleneck: identifying an object, writing an accurate description, setting a defensible price. Hours of manual research per sale, and no easy path to global buyers.

The AI that matters

A single vision model looks at a photo and says "round metal pin." Not sellable. The pipeline chains several models so each does the job it is best at, and grounds the result in live web search rather than in the model's memory - so a price arrives with real sources behind it.

"Round metal pin" becomes "1950s Magen David Adom Healthcare Association pin, ~$94, sourced from live eBay listings."

It runs at roughly $0.055 an item, asynchronously, so the item exists immediately and the enrichment lands a moment later. Every model and prompt change ships on measured win-rate from a blind A/B harness, not on how the output feels.

The call that shaped everything

The defining decision was go-to-market, not tech: integrate with the marketplaces buyers already use, before building a marketplace of my own. eBay's Sell API is open and reaches roughly 132M buyers, so meeting sellers where the demand already is beat asking them to bring their own audience. Integration-first then shaped the entire roadmap - one catalogue syndicating out to eBay, BidSpirit and Etsy, including one marketplace whose API had to be reverse-engineered from its own UI.

Production discipline

Multi-tenant from day one and hardened for real use: per-tenant isolation, encrypted marketplace tokens, OAuth across three providers, the full eBay onboarding gauntlet cleared, and event logging behind a monitoring dashboard. When two production outages hit in a single day, each was reverted in about three minutes - and the build now hard-fails on the class of error that caused the first one, so it cannot recur.

Outcome

A live multi-tenant SaaS with a real auction house onboarded, a real listing on eBay, a measured 79-84% unit margin, and a deploy pipeline hardened by real incidents - about 40 features shipped solo in four months.

What it demonstrates: the full spectrum one person is rarely trusted with at once - the strategic call, measured AI depth, the full-stack build, and the commercial machine to take it to market.