800,000 Parts, 10 Million Images: Building a Dataset for the Supercar Parts Market
For the past six months, I’ve been building a dataset around exotic cars and their parts: listings, prices, images, vehicle catalogs, OEM catalogs, fitment data, price history, and much more. Next up: Europe, MENA, China, and eventually the rest of the world. The goal is ambitious — build a global dataset of exotic car parts.
My previous post about researching this market got a great response, so I’ll keep sharing what I’m learning. In AI, algorithms are becoming commodities fast. A competitor can replicate a good feature. Replicating a proprietary dataset built from millions of real-world data points is a very different game. That’s why, for the past six months, I haven’t just been building the product — I’ve been building the data moat behind it.
The interesting part starts after collection. I turn every listing into structured data: vehicle, OEM number, fitment, condition, price. I track when a part appears, how its price changes, and when the listing disappears. Over time, this becomes more than a parts catalog — it becomes a history of how individual parts move through the market.
Images are even more interesting. A seller might write three words in the description, but the part itself contains much more information. OEM numbers, manufacturer numbers, markings, specifications, materials — all of this can be extracted directly from images and connected to the rest of the dataset.
And this is where it gets interesting.
At this scale, I can start identifying parts even when the seller has no idea what they’re selling. A few photos and minimal text should be enough to determine what the part is, identify its OEM number, and figure out which vehicles it fits.
Then there’s pricing intelligence. I don’t just know listing prices — I know how long individual parts stay on the market. So instead of answering “What is this part worth?”, I can answer a much more useful question: What should I price this part at if I want to sell it in a week, a month, or I’m willing to wait six months?
But the main problem I’m working toward sounds ridiculously simple:
Take a photo of a part → identify its OEM number.
Once that works, almost the entire selling process can be automated around it. AI identifies the part and fitment, analyzes the market, recommends a price, and generates a ready-to-publish listing.
Ideally, the seller doesn’t even need to know what’s sitting on the shelf. Take a few photos → get a part ready for sale.
That’s why I need 800,000 parts and 10 million images.
And I’m just getting started.
Images are even more interesting. A seller might write three words in the description, but the part itself contains much more information. OEM numbers, manufacturer numbers, markings, specifications, materials — all of this can be extracted directly from images and connected to the rest of the dataset.
And this is where it gets interesting.
At this scale, I can start identifying parts even when the seller has no idea what they’re selling. A few photos and minimal text should be enough to determine what the part is, identify its OEM number, and figure out which vehicles it fits.
Then there’s pricing intelligence. I don’t just know listing prices — I know how long individual parts stay on the market. So instead of answering “What is this part worth?”, I can answer a much more useful question: What should I price this part at if I want to sell it in a week, a month, or I’m willing to wait six months?
But the main problem I’m working toward sounds ridiculously simple:
Take a photo of a part → identify its OEM number.
Once that works, almost the entire selling process can be automated around it. AI identifies the part and fitment, analyzes the market, recommends a price, and generates a ready-to-publish listing.
Ideally, the seller doesn’t even need to know what’s sitting on the shelf. Take a few photos → get a part ready for sale.
That’s why I need 800,000 parts and 10 million images.
And I’m just getting started.
October 7, 2026
