Some of the most interesting AI problems aren’t in regulated finance at all — they’re in messy, high-variance consumer markets where a human currently squints at photos and does mental arithmetic. We built one as an internal capability demo: grade a collectible trading card from its listing photos, then model the economics of reselling it.
The manual version is tedious and error-prone. A buyer scans marketplace listings for ungraded cards, eyeballs each photo to guess what condition grade it might earn, looks up what that card sells for at each grade, subtracts grading fees, shipping and marketplace commission, and decides whether the margin is worth the risk. Do that across dozens of live listings and a single one-grade misjudgment can wipe out the profit.
A vision model with an opinion, and a rubric to hold it to
From a single search, the pipeline runs end to end per listing:
- Find — pull live marketplace listings for ungraded cards, automatically excluding already-graded ones.
- Grade — a vision LLM assesses condition from the listing photos against a standard four-pillar grading rubric — centering, corners, edges, surface — returning a single 1–10 grade and a confidence.
- Price — look up ungraded and per-grade market comps and recent sales.
- Model — compute the full economics: purchase price + grading tier + inbound/outbound shipping + marketplace fee, against an interpolated graded resale value → profit, margin, ROI, a risk score, and a buy/hold/pass call.
Results render as a sortable grid — rank by profit, ROI, confidence or grade — so a human sees the whole opportunity set at a glance instead of one tab at a time.
Why it’s a good stress-test
Two things make this harder than it looks, and that’s the point. First, the vision judgement has to be honest — a grade without a confidence is worthless when a single point changes the maths, so the model reports both and the financial layer treats low confidence as risk. Second, the data is hostile: listings are inconsistent, comps are noisy, sources rate-limit and fail. So the pipeline runs a resilient provider cascade with retries, fallback and defensive parsing of model output — the same fault-tolerance discipline we bring to production systems.
It’s a demo, not a product — no live deployment. But it’s a compact showcase of the pattern we build for clients all the time: a vision LLM that renders a judgement, fused with multi-source data and a real financial model, with confidence and risk carried all the way through.
The interesting part was never the cards. It was teaching a model to be usefully unsure.