The system of record
for your build.

Every DIY repair starts with identification, and the whole aftermarket is priced on the assumption that you already know what you're looking at. So the first bet was simple: make the camera the search box. I trained a vision model, built the scanner, and it worked.

Then I used it on my own car for a while and found the thing I had missed. Naming a part takes a few seconds and then it's done with. Running a build goes on for years: what is installed, what it cost, what the car puts down, when the fluid was last flushed, who else runs the same setup. Nothing was tracking any of that except spreadsheets and a shoebox of receipts. So the scanner turned into a feature and the build log turned into the product.

Fiddle's look is a system I call Holo: deep blue-black surfaces that lighten as they come toward you, thin strokes, light type, and teal saved for the one thing that matters. Eighty percent quiet, twenty percent glow. I tested the candidate palettes on the actual phone, in the actual garage, and kept the winner.

The restyle never stopped the product. The colors live in one tokens file, old names point at new values, and screens migrated as I touched them. The name took a few tries too: Partsnar, then MotorMuse, then fiddle. It's what you do with a car on a Saturday.

The detector is a YOLOv8 model I trained on 23 exterior part classes, exported to a 12.7 MB ONNX file that runs entirely in the browser. Instant scans, works on garage Wi-Fi, and no photo of your car ever leaves the phone. A scene classifier routes everything else (engine bays, interiors, parts on a bench) to Claude vision.

Owning the pipeline pays off in unglamorous ways. The model once labeled both mirrors “right,” so now paired parts get their side reassigned by position. Scanned parts are checked against product photos before the app will claim a match. None of that shows up in a demo, and most of it is what separates a demo from something a person will keep using.

The aftermarket has no clean API to buy, so Fiddle's catalog is one I went and built. 245,471 parts across 362 brands, with translated descriptions, harvested photos, and search fast enough to feel local. That number is smaller than what the crawl pulled down. A good deal of what comes back is fluff, duplicate listings and false positives, and the cleaning pass is what gets it to something you can search. On top of the catalog sits fitment: chassis-code dictionaries across dozens of makes, with an LLM filling in only where the dictionary abstains.

It's the least visible work in the product and the hardest to copy. Calling a vision API takes an afternoon. Spending months cleaning part records is the part most people skip, which is why it's worth doing.

The power readout shows your factory rating until a dyno sheet or a rated tune says otherwise. The fitment simulator flags, in amber, when its answer isn't about your car. Even sign-in refuses to leak who has an account.

From the outside those look like engineering choices. They're design decisions, and they come from the same confirmation-loop thinking I ship at Oracle, running in a garage app instead of a supply chain.

Fiddle is a two-person team where one of us is an AI. I bring the design judgment and the product decisions; Claude brings speed and a bottomless appetite for the grind. I review what it writes, and I say no a lot. The repo shows the partnership plainly: merged PRs from Claude's branches, and a 41-suite test bed that stays green.

When trying an idea costs an afternoon instead of a sprint, you stop arguing about mockups. The live AR overlay didn't die in a critique. It died in my driveway, jittering, where there was nothing left to debate. What becomes scarce is the quality of the decision, and that happens to be the thing a design career builds.

It's also the cheapest technical empathy I've found. When I sit down with data scientists and engineers at work, I don't have to imagine what a model can and can't carry inside a design. I ship against one every week, and it shows up mostly in which fights I decide are worth having.