One thing that kept bothering me about the pricing results was that two listings could look very different even if the text search thought they were a match. So I added a local vision LLM step that looks at the actual images and scores how well an eBay listing visually matches the pin you uploaded. This turned out to be one of the most satisfying upgrades so far.
Instead of mostly sorting results by price and hoping the good matches float up, the system can now put the visually closest listings first. That makes the whole pricing view feel much more grounded in reality, especially for pins with similar names, multiple versions, or confusing seller descriptions.
It also runs entirely on the mini server at home, which I love. No extra cloud API bill, no per-image charges, just local hardware doing the work. There is something very fun about squeezing this much usefulness out of a small box in the house.