Two dealers supply the watches — one through an online store and a live availability feed, one through spreadsheets that change shape between deliveries. This pipeline brings both together, hosts every photo on a CDN, has an LLM write a constrained auction listing for each watch, and exports an upload-ready catalogue.
A small codebase with no web interface — operators run it from the command line and a scheduler keeps it current. The interesting work is not the size: it is making unreliable inputs and a probabilistic writer produce output an auction house can publish.
The dealer’s store carries descriptions, specs and photos, but its stock status lags. A separate feed is authoritative for what is still for sale — and a lot for a watch already sold is a problem at the auction.
Between two deliveries the file lost five columns, gained three, renamed four and moved its images into differently named columns.
Every lot needs a fixed-format title, a structured description with specifications, a condition report, a reserve, estimates, and IDs from the platform’s taxonomy.
Copy that calls a dial “sunray” or a part “original” when the data does not say so is a liability, not a style choice.
Merge sources it does not control, track availability close to real time, write hundreds of listings without inventing facts — and never pay to rewrite a listing that has not changed.
Scrapes the dealer’s storefront twice a day, polls its availability feed every 30 minutes, applies negotiated price overrides, and imports wholesale spreadsheets on delivery.
An LLM writes the title, a four-paragraph description, a specifications list and taxonomy selections for every eligible watch.
Downloads every photo, normalises its format and hosts it on a CDN.
Exports upload-ready lots with reserve, estimates and an approved condition report.
Emails a list of watches that have left the feed, with photos and prices.
Compares two catalogues so a new export can be reviewed as a set of changes.
The store is scraped for its content, but a watch only reaches the catalogue while the feed still lists it. The store lags; the next export after a sale simply does not contain that lot.
Listings started out tied to the storefront. When the wholesaler arrived, a listing became something either source can own — migrated in place, with every existing listing carried over.
A watch missing from a scrape or delivery is retired, not deleted. One that comes back keeps its lot reference and listing, and the history of what sold stays visible.
Headers are normalised before matching, so a reordered file, a renamed column or a stray space in a header does not break an import. Irrelevant columns are ignored explicitly, so a warning about an unrecognised one still means something.
The importer accepts other names for the wholesale price — but never a retail or MSRP column, which would silently become the auction reserve. Using one has to be an explicit choice.
Every row is validated before the first write; one bad row rejects the delivery as a whole, with the errors listed. A dry run shows what would be created, updated and retired.
Watches missing from a delivery are retired — but never after a partial import, which would retire real stock.
“A partial import is not a valid inventory snapshot, so never reconcile against one.”
Each watch carries a fingerprint of the facts that shape its listing. When it has not changed, the model is not called. Prices are deliberately left out, so a price change updates the catalogue without paying for new copy.
Adding a field is the trap: it changes every fingerprint and would regenerate every listing. New fields therefore count only when a watch actually has a value for them.
when a delivery introduced new fields — the fingerprint rule kept every existing listing as it was.
The model returns structured output against a schema — title, description and taxonomy selections, all required. The constraints come in three layers.
Fixes the title format and length, the description structure and a required specifications list; translates the dealer’s shorthand; keeps reference numbers exactly as given; bans unsupported descriptors; and embeds the platform’s 636-ID taxonomy with an instruction to never guess a brand.
Prices and internal identifiers are removed before prompting. The model cannot mention a price it was never given.
Every response is validated before it is saved: the specifications list must be present and complete, and the listing must carry the watch’s reference. A failure is retried with a note naming what was missing — then the job fails rather than store an incomplete listing.
Separate tooling finds and regenerates any listing written before a rule existed, and every listing records which model wrote it — so a later quality question can be traced to the model that caused it.
One supplier served WebP images under .jpg addresses while declaring them JPEG. Neither the address nor the declared type could be trusted, so the format is read from the file itself.
JPEG and PNG are stored untouched, since re-encoding a JPEG only loses quality. Other formats become JPEG, flattened onto white so transparency does not turn black.
A failed re-download keeps what was already hosted, and a missing image is left visibly missing so the next run retries it.
Watches whose photos are all hosted are skipped entirely.
The model, network, storage and queue are all faked — so the suite runs offline, costs nothing, and can script a model that returns exactly the broken output a test needs.
A delivery arrived with a fifth of its columns removed, added or renamed. One new column looked like a merge of two old ones; the data said otherwise — in 44 of 45 rows holding both, the values were identical, and the one difference was the newer value being more specific. It was a rename, and treating it as one avoided duplicated data. The delivery imported 627 watches and produced 461 new listings in one pass.
Some generated descriptions ended before their specifications. The fix was layered: a prompt rule, validation with retry, a hard failure instead of a stored partial listing, and a repair pass over the listings already affected.
Files served with the wrong extension and type were stored wrongly until detection moved to the file contents.
Repeated phrases, unworn watches receiving the pre-owned condition report, and flourishes the data did not support were each closed with a rule and, where possible, a test.
An auction lot for a sold watch is worse than a missing one.
Suppliers reorder and rename columns; position is the least stable thing in the file.
A retail figure would silently become the reserve.
A half-imported delivery is harder to reason about than a rejected one.
Retiring against a partial file would retire real stock.
A price change should update the catalogue, not pay for new copy.
The model cannot state what it was never given.
An incomplete listing saved is worse than a job that visibly failed.
The supplier’s addresses and declared types were wrong.
A returning watch keeps its lot reference and history.
Three inputs from two dealers merged into one list of what is for sale, refreshed every 30 minutes.
1,027 AI-written listings with specifications and taxonomy selections.
3,669 photos normalised and hosted on a CDN.
27 upload-ready exports of up to 663 lots, with pricing, estimates and condition reports.
I designed, built and coded the pipeline: how two unrelated sources map onto one listing, the rules for what a malformed spreadsheet may change, the image pipeline, the prompt and the checks between the model and a published listing, and the export the auction platform consumes. I did the coding myself, working with AI-assisted tooling. When a delivery behaved unexpectedly, I traced each anomaly to a root cause with the numbers to back it, and wrote it up for the client.
Built at Serenity Gallerie, an online auction house for luxury jewellery and watches, where I have been Tech Lead since Q4 2025. Dealer, supplier, platform and domain identifiers are deliberately omitted.
Shipped in 2025. Staff upload a file and review finished lots, and the model is never allowed to invent stone counts, carats or metal purity.
Read the case study →SaaS · Cloud marketplace · Server automationv1 shipped February 2025. Sites, SSL, PHP, databases, backups and WordPress itself are run from the browser, with Stripe billing on top.
Read the case study →Next