Estimonia: AI Photo Valuation for Antiques

Most people just want to know if it's worth anything

Most people holding an inherited watch, clock or ring don't need a certified appraisal. They need a first answer: is this worth a specialist's time, or not? Appraisers are slow and expensive for that question. A normal image search shows listings, not a value.

An AI answer alone isn't a product

A vision model (AI that reads images) will happily name an object and quote a price. But the answer has to follow a fixed format, stay within set limits and be honest about uncertainty. It also has to handle blurry photos, messy backgrounds and things that aren't collectibles at all. And failed calls can't quietly cost money each time.

Nobody wants to sign up just to ask

Most people use it once and care about privacy. A sign-up wall before the first valuation would lose most of them. So the system had to work without an account, while still letting people keep what they scanned.

One API behind the scanner

A Laravel 12 API (PHP 8.4, MySQL) powers the web scanner. The API is versioned and gives out anonymous Sanctum tokens, so a first scan needs no account. Anonymous users can create an account later without losing their items.

Valuations run in the background

Each upload of up to four photos creates a Scan. A background job (ProcessScanJob, run on Redis and Laravel Horizon) sends the images through Prism to OpenAI GPT-4.1 vision. The AI must reply in a strict JSON format. The reply includes a verdict on image quality with warning codes, the item's name and model reference, and its materials. It also gives a tight USD min/max range, confidence and rarity scores, and an explanation based on similar items. The job stays within the provider's rate limits. When a call fails, it tries again with longer waits each time. It then saves the result as an Item. Uploads are resized to 2000px before they are stored.

A clear view of cost and failures

Every AI call is logged with its token use, cost and outcome (success, rejected or failed). Spend is recorded before the reply is read, so broken replies still count. An admin dashboard shows daily spend, cost per scan, a budget forecast and money spent on rejected photos. Limits per IP address and per user protect the parts that work without an account.

Reuse instead of new pipelines

Later features run on the same Scan → Job → Item path instead of copying it. The estate-valuation flow uses it: it values a whole household from a batch of photos, shared through a private link. The opt-in introductions of sellers to auction houses use it too. Cost logging, value tiers and retries apply to them automatically.