This summer I went to look at six versions of essentially the same used car.
It had all looked so easy on Cars.com at my desk. Twenty minutes of searching produced a shortlist, a map, a plan for the day. Then I drove to the first lot.
By evening I had twenty photographs and six videos buried in my camera roll, plus six packets of dealer paper: Carfax reports, service histories, sales sheets, documentation fees, add-ons, paint protection, registration charges. Each packet was formatted just differently enough that nothing compared cleanly. Which one had the new tires? Which seller said the brakes were done? Did the $899 fee belong to the black car or the blue one?
The search took twenty minutes. The looking took all weekend and gave me a headache.
Somewhere between the third lot and the fourth folder, the obvious thing became clear. The tools are excellent at finding cars. The actual decision happens in a driveway or a parking lot, with your phone in one hand and somebody else's paperwork in the other.
AI solved abundance before anyone solved memory.
So I built the app I wanted in those parking lots.
The idea was narrow on purpose. Not another chatbot. Not a mechanic. A dossier: one durable record per car, made from what I actually saw. Photograph the VIN. Photograph the tires, the window sticker, the worksheet. Hold the mic and say what I noticed. Let the software read the paperwork, recompute the real out-the-door price, and flag what doesn't add up. Then compare cars from evidence instead of recollection.
The next weekend I went shopping again with a beta on my phone and another list of six. I saw one. By the time I'd walked around it, photographed the paperwork, and read the app's briefing, I didn't need the other five. I bought it that morning. With confidence.
Then the second half of the idea took over. Buying a car is a weekend. Owning one is years. The same record that answered "should I buy this one?" was suddenly the record that would answer "what has happened to my car?" and, eventually, "why is this car worth more?" So one button, I bought it, turns the candidate into the car. The inspection becomes the acquisition record. The odometer becomes baseline mileage. From there, every receipt gets photographed and filed. The glove box moves onto the phone.
A few things I decided early and haven't regretted.
Your data stays on your device. There is no account, no ad model, and no reason for me to hold your vehicle history in a database somewhere. When a remote model reads a photographed document, it's transient work. The answer comes back to your phone and the model forgets.
AI reads. The app verifies. You decide. Multimodal models are remarkable at reading a crumpled dealer worksheet. They are not to be trusted with arithmetic. So the deterministic questions, VIN check digits, totals, fee classification, mileage comparisons, are handled by ordinary code, and where the app isn't sure it says so. A missing number is better than a wrong number.
No subscription. You shouldn't rent the history of something you already own. Shopping is free. Ownership will carry a one-time price, and the people who downloaded the first release will never see it.
DriveDossier is made by Ultra-Normal, the small studio where I build tools for people who do the slow work. It's the first one that started in a parking lot. I hope it saves you the headache.
Justin Neuman
New Haven, September 2026